memorix

Changelog

All notable changes to this project will be documented in this file.

Unreleased

Fixed

• Fixed consolidation false positives for Korean and punctuation-only records; queued automatic cleanup jobs now honor auto_cleanup = false at execution time (#385).

• Observation writes now persist an idempotent CodeGraph/evidence projection job in the same SQLite transaction; successful foreground projections acknowledge it, while failures remain recoverable by maintenance.

• Added common DerivedView adapters for SQLite FTS5 and Orama with explicit status, rebuild/refresh, compact, and close lifecycle methods.

• Added a CodeGraph DerivedView adapter for Lite refresh, status, SQLite compaction, and lifecycle shutdown.

• Release workflow now runs automatically for published GitHub Releases and requires npm visibility before MCP Registry publication.

• Workset assembly now accepts only request-scoped derived-view receipts, preventing a concurrent search from leaking FTS/Orama state into context.

Changed

• Search execution now returns a request-scoped receipt to compact search, index probes, and TUI chat, so concurrent requests cannot exchange diagnostics.

• Observation inserts and topic-key updates commit canonical SQLite rows with vector, qualification, and claim-derivation jobs in one transaction.

• A deduplicated maintenance job records writes received while it is running and schedules a follow-up pass, including after worker failure or lease loss.

• Maintenance progress now has one kernel result contract; negative, fractional, non-finite, and inconsistent counts cannot be persisted as successful work.

• SDK, TUI, compact search, and MCP retrieval now share the RetrievalFacade; feedback state is read from the active canonical store and reported separately from embedding fallback.

• Topic-key updates publish their Orama projection only after the canonical transaction commits; failed updates or linked-record callbacks leave the prior searchable record intact.

Verification

• The 20-case context task matrix and the 40,000-record local store gate pass on Windows/Node 22 with embeddings disabled. This is offline fixture evidence, not a claim about remote embedding quality.

[1.10.3] - 2026-11-05

Added

• Pi 1.0+ native MCP setup through project or user mcp.json, with stdio and HTTP transports alongside the existing extension and shared skills (#367).

Fixed

• CodeGraph now retains two snapshots, sweeps stale derived symbols and edges, compacts SQLite after cleanup, and supports codegraph.auto_refresh.

• Lazy large-corpus mode now initializes vector schema and backfills missing vectors from durable SQLite/Lance state (#374).

• CJK-heavy search avoids fuzzy traversal over character and bigram tokens; project affinity is bounded to a tie-break so strong lexical matches stay on top (#373, #375).

• Dashboard and TUI browser launches no longer interpolate URLs into shell commands; release workflows retain the existing tested action references.

• Runcrate memory preset -- optional background memory LLM processing through an OpenAI-compatible endpoint for open-weight models, with an isolated RUNCRATE_API_KEY, CLI/TUI setup and bilingual configuration guidance.

[1.10.2] - 2026-10-30

Fixed

• LanceDB semantic-index writes are batched and periodically optimized, with memorix repair index available to compact older inflated indexes safely. Startup now checks persisted vector IDs before scheduling re-embedding or rewriting cached vectors.

• Derived-view receipts now report degraded embedding health and redact local paths and credentials from failure reasons.

• Hermes now registers bundled skills by file path, matching the current plugin API (#366).

• MCP search results and the session-start card are available in structuredContent for clients that do not display text blocks (#368).

[1.10.1] - 2026-10-23

Added

• Cheaper Inference memory preset -- optional background memory LLM processing through an OpenAI-compatible endpoint, with an isolated CHEAPER_INFERENCE_API_KEY, CLI/TUI setup and bilingual configuration guidance.

• Derived-view receipts -- context and search receipts now expose privacy-safe FTS, Orama, embedding, and CodeGraph freshness, generation, rebuild status, and bounded failure reasons.

[1.10.0] - 2026-10-23

Added

• Measured context evaluation -- adds a reproducible 20-case runtime matrix covering continuity, retrieval, safety, runtime binding, and integration parity, with token, latency, RSS, fallback, degraded, and review-cost metrics.

• Large-store release gates -- records and verifies 40k and 100k local capacity baselines across SDK, HTTP MCP, hooks, SQLite persistence, and SDK reopen paths.

Changed

• Retrieval receipts -- compact search, MCP search, and CLI JSON search now expose request-level candidate, selected, omitted, token-budget, and fallback diagnostics while keeping query text, raw memory content, and local paths out of the receipt.

• 1.10 development contract -- documents memory ownership, runtime binding, adapter parity, and the measured context baseline as maintained acceptance surfaces.

[1.9.12] - 2026-10-17

Fixed

• Release provenance gate -- the publish workflow now verifies npm registry signatures and fails closed when provenance publication cannot complete; it no longer retries with an unprovenance --ignore-scripts publish.

• Background update installs -- opt-in automatic updates disable npm lifecycle scripts during the global install.

[1.9.11] - 2026-10-16

Fixed

• CJK full-text search -- the in-memory Orama index now keeps the default Latin tokenizer behavior while adding bounded character n-grams for Chinese, Japanese, and Korean text, so exact titles and short CJK queries work when embeddings are disabled.

• Runtime supply-chain baseline -- refreshed the MCP SDK, Git adapter, HTTP client, URL/IP parsers, Hono, LanceDB, Sharp, and related lockfile resolutions; the production dependency audit is now clean.

• Large-store alias migration -- startup project-ID normalization now uses a bounded SQLite update instead of requiring the full observation corpus in memory. The FTS5 triggers and storage generation stay current.

• Alias-aware search visibility -- observations retained under a merged project alias are readable through the same alias-expanded search scope.

• Hybrid search recovery -- an empty or stale derived semantic result now falls back to the durable lexical path instead of becoming a silent empty answer.

• Search CLI parity -- memorix search "query" and memorix search --query "query" now share the same argument contract.

• Pi hooks on Node 26 -- hook stdin is paused and unrefed without closing file descriptor 0, so subsequent synchronous Git detection remains safe when a host supplies hook input through a regular file.

Added

• Qwen Code hooks -- memorix hooks install --agent qwen writes the official .qwen/settings.json hook shape, preserves user settings, and installs project/global QWEN.md guidance. iFlow is not targeted because its official CLI service was retired in April 2026.

• API Route memory preset -- optional background memory LLM processing through an OpenAI-compatible endpoint, with an isolated API_ROUTE_API_KEY, CLI/TUI setup and bilingual configuration guidance.

[1.9.10] - 2026-10-08

Fixed

• OpenCode 1.x regression -- generated plugins now include the v1 server field alongside the v2 setup contract, so OpenCode 1.x can load the legacy named export. Lifecycle, file, command, message, and compaction events are also forwarded through the v1 generic event hook, with a longer Windows hook timeout to reduce stale Bun timer failures.

[1.9.9] - 2026-10-02

Added

• Requesty memory preset -- memorix init and the configuration TUI can select Requesty for optional background memory LLM work through its OpenAI-compatible endpoint, with provider-specific key isolation. REQUESTY_API_KEY is also accepted for Requesty embedding endpoints.

[1.9.8] - 2026-10-02

Fixed

• OpenCode v2 plugin compatibility -- generated plugins retain the v1 named export while also exposing the v2 default id/setup contract, including event subscriptions, tool hooks, cleanup, and host session IDs.

[1.9.7] - 2026-09-28

Fixed

• SQLite FTS5 startup resilience -- transient SQLITE_BUSY contention during lexical-index initialization is retried before FTS5 is marked unavailable.

• Large-store search fallback -- runtime embedding failure, provider disappearance, and vector-dimension mismatch now all fall back to the durable SQLite lexical index instead of returning a silent empty result.

• Release metadata -- synchronized the npm, MCP Registry, and packaged integration metadata for the 1.9.7 patch release.

[1.9.6] - 2026-09-22

Added

• Task continuity ledger -- adds a project-scoped append-only ledger for requirements, decisions, verification, risks, and outcomes. The bounded Project Context can include a ledger by taskId; no automatic promotion or LLM summarization is performed.

• Continuity outcome projection -- adds deterministic validated, at-risk, in-progress, and unverified states plus replay coverage. Diagnostic scores are exposed in receipts only and do not silently reorder memories.

• Infrastructure lifecycle closeout -- bounds HTTP sessions and SQLite handles, reuses project business runtimes behind fresh modern MCP protocol shells, records event-loop/resource health, and adds a deterministic memory quality gate.

• Atlas Cloud memory preset -- memorix init and the configuration TUI can select Atlas Cloud for optional background memory LLM work through its OpenAI-compatible endpoint, with provider-specific key isolation.

Fixed

• Configuration scope isolation -- explicit home/project resolution now applies consistently to TOML, YAML, legacy JSON, and .env loading, so long-lived processes and tests cannot reuse another configuration scope.

• OrcaRouter documentation -- documents the existing OpenAI-compatible gateway path with fixed-model and fallback guidance; no vendor-specific provider branch is required.

• Capacity closeout -- TeamStore queries, maintenance history, embedding caches, HTTP project caches, and large-store behavior are bounded and regression-covered.

• Stable coordination queue order -- available tasks now dispatch FIFO with a SQLite rowid tie-breaker, preventing cross-platform adapter-routing flakes when task timestamps collide.

• Continuity correctness -- verification obligations can move between statuses by stable ID, retries are idempotent, open risks are visible in the outcome state, and the 200-event cap is enforced transactionally.

• Replay-safe orchestration evidence -- tool effects are recorded with redacted fingerprints, risk tiers, replay policy, and unknown-after-crash status without persisting raw tool arguments or output.

[1.9.5] - 2026-09-20

Added

• Setup HTTP MCP -- memorix setup --mcp http honors Grok (writes a [mcp_servers.memorix] URL in Grok config.toml) and rewrites installed plugin/extension bundled MCP manifests from memorix serve to http://localhost:3211/mcp. Default Grok MCP stays host-owned. Docs include generic systemd user unit and LaunchAgent KeepAlive examples for memorix serve-http.

• User-global store sync -- memorix sync store push|pull|status --scope user replicates every local project's eligible observations through the existing opt-in relay. Default --scope project is unchanged. Privacy filters still drop personal, agent-targeted, candidate, and ephemeral rows. The live SQLite file is never copied.

Fixed

• HTTP control-plane safety -- JSON request bodies are bounded, non-loopback binds require an explicit Bearer token or unsafe opt-in, listen failures are reported synchronously, background startup removes unverified state, and shutdown waits for HTTP and SQLite resources to close.

• Project-scoped graph storage -- graph entities and relations now carry a project key, legacy JSONL migration recovers ownership from observation references, and Dashboard/MCP graph access no longer crosses projects.

• Migration and lifecycle reliability -- legacy observation/subdirectory migrations use locks and atomic writes, malformed observation JSON fails closed, stale file locks cannot evict a live writer, stateless MCP bindings expire and are cleaned up, and SDK clients share resources safely.

• User-scope sync correctness -- imported observation IDs are remapped on collision, tombstones validate their source project, and user namespaces must be explicit per-user values.

• Release contract -- publish workflow validates tag/version/commit identity, waits for npm visibility before MCP Registry publication, and keeps prepublish checks non-mutating.

[1.9.4] - 2026-09-18

Added

• MiniMax first-frame video input -- image-to-video generation now accepts a public HTTPS first-frame image URL through the CLI and MCP surfaces, preserves it in durable media jobs, and submits it with the provider's first_frame image role.

Fixed

• First-frame URL safety -- rejects non-HTTPS URLs and embedded credentials before a media request is submitted or persisted.

Changed

• Grok Build documentation -- the README comparison now uses the official Grok Build logo assets.

[1.9.3] - 2026-09-14

Fixed

• SQLite storage reliability -- transient cross-process locks no longer turn a process into a permanently empty read-only store. Observations, sessions, and mini-skills share the same initialization and retry contract; persistent runtime failures remain visible.

• Storage diagnostics -- memorix doctor reports the active backend, distinguishes unavailable storage from an empty project, and returns a failing exit status for storage failures.

[1.9.2] - 2026-09-11

Fixed

• Non-interactive MCP startup -- running the package without arguments through piped stdio now starts the MCP server, while a real terminal keeps opening the bundled memcode agent. This makes generic npm MCP launchers and registry handshake checks use the correct protocol entry point.

• Maintenance initialization -- dashboard Consolidate uses its request-scoped observation store, and CLI deduplication initializes the project memory LLM before checking availability.

[1.9.1] - 2026-09-08

Changed

• Large-scale retrieval -- SQLite FTS5 is now the persistent lexical candidate index. It stays synchronized on insert, update, delete, and can be rebuilt from the durable observations table.

• Disk-backed semantic retrieval -- the optional local LanceDB shadow index stores only vector IDs and metadata, uses HNSW with scalar quantization once the corpus is large enough, and fuses bounded semantic candidates with FTS5 results. Search falls back to Orama when the optional native package is not available.

• Lazy large-corpus startup -- projects above the Orama hydration threshold keep SQLite as the source of truth and do not load the complete observation corpus into the interactive agent process. Detail, export, and maintenance paths can still load the data they explicitly need.

• Large-store acceptance -- adds regression coverage and verified gates for 1k, 10k, 40k, and 100k records without reducing the durable memory corpus.

• Post-threshold vector writes -- newly generated vectors still reach the persistent semantic index after the in-memory Orama corpus is released.

Fixed

• Stale lexical index on upsert -- replace-style SQLite writes could leave old terms in an external-content FTS index. Observation upserts now preserve trigger-visible row updates so old terms are removed.

• Optional native dependency bundling -- LanceDB is resolved at runtime so unsupported platforms or installations without its native binary continue to build and use the FTS5/Orama fallback.

[1.9.0] - 2026-09-07

Added

• Opt-in multi-device sync -- adds a project-scoped observation event contract with privacy filtering, local outbox retry, conflict evidence, device-clone protection, bounded relay pulls, and filesystem/GitHub/S3/ Postgres transport boundaries. It is disabled by default and keeps local SQLite as the canonical store.

• Sync hardening -- retracts records that become private or agent-targeted, keeps imported observations and sync state atomic, rejects immutable-key collisions, verifies every pulled content hash, and uses stable continuation cursors so large pulls do not skip events.

• Resource hardening -- coalesces deferred vector workers across hook processes, respects retry cooldowns, drains due vector jobs through one recoverable worker, removes the redundant full-document detail cache, and batches startup vector hydration to reduce CPU/RAM peaks without imposing a small corpus limit.

[1.8.9] - 2026-09-06

Added

• Grok Build integration -- adds native lifecycle hook capture for session, prompt, tool, compaction, and stop events, plus Grok-specific AGENTS.md guidance. Thanks to @RaviTharuma for the original integration.

Changed

• Host-owned Grok MCP -- memorix setup --agent grok --global uses Grok's hook and rules surfaces without writing Grok's config.toml or MCP entries.

• Agent documentation -- the English and Chinese README matrices and setup guides now list Grok Build and its trust model.

Fixed

• Safe Grok hook updates -- preserves unrelated user hooks, removes only Memorix-owned entries, supports GROK_HOME, and reports project hook trust accurately without claiming that Memorix can approve it.

[1.8.8] - 2026-09-03

Changed

• Public product website metadata -- package, MCP Registry, Agent plugin, installer, README, and GitHub repository links now point to the live Memorix website at https://mem.rglens.com. This patch does not change the runtime behavior of the memory layer.

[1.8.7] - 2026-09-02

Added

• CodeGraph semantic foundation -- adds a TypeScript Compiler API provider for source-backed TypeScript, TSX, and JavaScript symbols, imports/exports, references, calls, inheritance, and test relations. Unsupported languages keep the bounded Lite structural fallback.

• Unified CodeGraph evidence -- persists fact provenance, parser errors, semantic coverage, affected import dependents, and bounded task context so agents can tell verified relations from heuristic hints.

• SCIP boundary and Dashboard health -- accepts bounded scip print --json outlines through the validated external-provider boundary and exposes read-only CodeGraph coverage in CLI doctor, HTTP, and the Dashboard without installing or taking over external indexes.

• Reproducible scale probe -- adds an opt-in generated fixture benchmark for measuring public CLI refresh and status latency under an explicit E: temporary directory.

Fixed

• Incremental semantic refresh -- re-indexes local import dependents when a changed file can alter their resolved relations, replaces each semantic slice atomically, and records a visible partial/failed state instead of claiming a ready graph after compiler failure.

• Code context precision -- unrelated symbols no longer enter semantic task context merely because they are functions, and freshness counts no longer duplicate file and symbol references for one memory.

• Modern HTTP project binding -- the stateless server/discover path now offers a reusable Memorix project-handle extension after verifying the Git root; legacy initialize sessions remain unchanged.

• Published runtime boundary -- TypeScript is a production dependency and remains external to the bundle so the semantic provider works after npm installation without inflating the runtime artifact.

[1.8.6] - 2026-09-01

Added

• WorkBuddy integration (#204) -- adds project and user-level MCP setup, project AGENTS.md guidance, safe no-op behavior for WorkBuddy's absent hooks surface, and uninstall discovery. Thanks to @wjxn13 for the original implementation and follow-up fixes; the contributor commits remain in the merge history.

Fixed

• Continuation intent (#266) -- a bare 继续, continue, or resume no longer triggers an expensive prior-work brief, while short substantive requests such as 继续修复, resume bug, and 继续 API still receive continuity context. Explicit handoff vocabulary remains supported. Thanks to @Tom-Ma-Ming for the original fix and regression cases.

• Star History assets -- refreshed the self-hosted chart data and light / dark SVG assets.

[1.8.5] - 2026-08-30

Fixed

• Modern MCP transport -- memorix serve and the HTTP control plane now use the official MCP v2 server path for 2026-07-28 discovery, request metadata, result envelopes, response metadata, stateless requests, and the subscriptions listen surface while forwarding business tools through the existing Memorix implementation.

• Background start race (#259) -- background startup now holds an exclusive, heartbeating lock across the check/spawn/readiness window. Lock contention fails closed with a retry message instead of starting an unlocked duplicate control plane; stale dead-owner locks are recoverable.

• Windows and EOF behavior -- stdio discovery remains compatible before and after readiness, roots-based project switching and deferred maintenance are preserved, and stdin EOF closes the official transport cleanly.

• Local build reliability -- the large runtime, CLI, and memcode bundles build sequentially, reducing Windows memory pressure without changing published entry points. Slow CLI uninstall tests now use asynchronous, hidden child processes so the full test run exits cleanly.

• Time-stable graph evaluation -- graph-context fixtures pin their reference date so retention policy changes do not make the test depend on the calendar.

MCP Compatibility Boundary

• Modern Tasks are explicitly unsupported because Memorix does not expose a durable task store for MCP task semantics.

• List responses expose a private zero-TTL protocol hint; Memorix does not claim a reusable server-side list cache.

• The external @hasmcp/mcp-spec-test runner currently fails on this Windows host before reaching the target because it imports drive-letter paths as ESM URLs. Production verification therefore includes official v2 direct tests, real stdio/HTTP smoke, and the legacy compatibility path; this external runner result is not reported as a pass.

[1.8.4] - 2026-08-27

Fixed

• Stateless HTTP MCP requests (#255) -- clients such as Claude Code that send the current protocol header on every request can list and call tools without echoing a custom project handle. Durable handles are still created on initialize and remain available when a client opts into them.

• MCP protocol self-description (#257) -- discovery and HTTP response headers now advertise the legacy revisions implemented by the pinned SDK, instead of claiming complete support for the newer 2026 revision.

• MCP stdio smoke contract -- the release smoke script now checks the truthful advertised revision set after the compatibility fix.

MCP Compatibility

• The 1.8.4 line supports legacy stateful MCP, versionless stdio tool calls, pre-initialize discovery, and the narrow stateless HTTP compatibility path. It does not claim full 2026 result metadata, Tasks, subscriptions, or list caching support.

[1.8.3] - 2026-08-25

Fixed

• Stdio MCP startup gate -- memorix serve now reads and bounds newline-delimited JSON-RPC input before project initialization completes. Early server/discover requests receive the complete discovery result, while initialize, notifications, responses, and versionless tool requests replay through the pinned SDK transport in input order without duplicate handling.

MCP Compatibility

• Stable cold-start discovery -- real stdio CLI startup now supports pre-initialize discovery within the startup window and preserves legacy initialize plus versionless tools/list/tools/call behavior. Tasks, subscriptions, list-response caching, and modern non-discovery result metadata remain unsupported or partial.

[1.8.2] - 2026-08-25

Fixed

• Dashboard deduplication initialization (#248) -- standalone and HTTP maintenance previews now idempotently resolve the selected project's memory LLM from TOML and .env, clear stale cross-project configuration, and keep credentials out of responses.

• CodeGraph capability fallback (#250) -- a failed or invalid context probe now tries the official explore CLI with bounded, shell-free spawning. Only validated paths, symbols, and relationships can select the external provider; separate context/explore diagnostics explain Lite fallback.

MCP Compatibility

• Discovery adapter (#249) -- the stdio server registers a real low-level server/discover handler that returns the required 2026-07-28 DiscoverResult before initialize. Legacy 2025 initialize and versionless tools/list/tools/call remain supported through the SDK's legacy envelopes. Tasks, subscriptions, and list-response caching remain unsupported and are not advertised.

[1.8.1] - 2026-08-24

Changed

• Truthful coordination status -- the persistent SQLite TeamStore is the sole coordination truth for agents, tasks, messages, locks, handoffs, and poll state. The Dashboard exposes read-only Coordination Status and Task Status views, and /api/team reports explicit error or degraded status instead of presenting a failed coordination read as a healthy empty result.

[1.8.0] - 2026-08-23

Added

• Persisted Evidence Cards -- observation provenance is stored in SQLite with project ownership, source reference, locator, captured hash, files, verification state, freshness, and audit events. Cards are available through memorix_evidence, memorix evidence, and the Dashboard Evidence Console.

• Auditable memory feedback -- verification, use, correction, source-change, conflict, strengthen, weaken, and revoke events materialize a bounded retrieval/retention weight without deleting the original source record.

• Stateless MCP project handles -- 2026-07-28-style requests can use an explicit Mcp-Project-Handle that survives service restart, while legacy stateful Streamable HTTP remains supported.

• One-shot CLI fallback budget -- repeated MCP-unavailable probing for the same project/task is deduplicated and fail-closed after one bounded context request.

• Evidence Console -- Dashboard includes a responsive, project-scoped evidence view with source, locator, freshness, verification, and audit count.

Changed

• HTTP TeamStore resolution is project-data-directory scoped, preventing A/B coordination state from crossing when sessions bind different projects.

• The 1.8.0 documentation records the shipped compatibility adapter, explicit non-goals, fallback semantics, and the Antigravity/Gemini setup matrix.

Fixed

• Preserve the 1.7 micro MCP profile's nine-tool budget by keeping Evidence and Feedback in lite/team/full profiles.

• Keep legacy relevance unchanged until a memory receives explicit feedback.

• Keep SQLite-backed HTTP helpers lazy so mocked and lightweight MCP startup paths do not open the database during module evaluation.

[1.7.3] - 2026-08-22

Added

• Optional HTTP rerank -- Thorough search with at least 3 candidates can reorder top hits through a configurable Cohere-compatible POST {base_url}/rerank endpoint ({model, query, documents}). Works with any compatible /rerank API. Configure [rerank] or MEMORIX_RERANK_*; base URL inherits the memory LLM lane when unset. Bearer inherits only when the effective rerank URL is that same trusted endpoint; a different host requires MEMORIX_RERANK_API_KEY / rerank.api_key. The model id is user-configured. When HTTP rerank is configured, a miss or timeout keeps the original order (no LLM fallback). LLM rerank remains available when HTTP rerank is off.

Fixed

• OpenAI-compatible provider base URLs -- preserve explicit versioned paths such as Volcano Ark's /api/v3 for ordinary chat, tool calling, streaming, and vision requests instead of appending an incorrect /v1 segment.

• HTTP rerank honors MEMORIX_RERANK_TIMEOUT_MS on the inner fetch (default 30s) instead of aborting at 5s.

• HTTP rerank no longer inherits or transmits the memory LLM credential when rerank.base_url points at a different endpoint. If a Memory LLM key is present, that foreign host is not called until MEMORIX_RERANK_API_KEY / rerank.api_key is set.

• Codegraph CLI tests close git/sqlite sandboxes with retries so Windows afterEach cannot hang past vitest's 10s hookTimeout.

[1.7.2] - 2026-08-19

Changed

• Faster large-store cold retrieval -- startup hydration now inserts persisted observations into Orama in bounded batches. This avoids rebuilding index balance once per row while continuing to yield to HTTP health traffic.

• Fail-closed large-store evidence -- the reproducible 40,000-record gate now exercises and times HTTP readiness, MCP initialize/bind/context/search/ store, hook capture, and SDK reopen behavior with hard request and release budgets instead of checking persistence alone.

• Isolated index tests -- startup hydration tests explicitly disable the embedding lane and restore the caller environment instead of reading an operator's provider configuration or local vector cache.

• Current public guidance -- README and canonical development/status docs now describe preview-first Dashboard maintenance and the 1.7 product line.

[1.7.1] - 2026-08-19

Fixed

• Contained graph text -- long graph node names, summaries, and relation targets now wrap inside the Dashboard tooltip and inspector instead of overflowing their containers.

• Centered maintenance preview -- the Cleanup preview dialog is centered on desktop and narrow screens, including after the Dashboard global reset styles are applied.

[1.7.0] - 2026-08-18

Added

• Preview-first Dashboard maintenance -- the Observations page now offers project-scoped Cleanup, Consolidate, and Deduplicate previews, while the Retention page can preview and archive canonical retention candidates. Every mutation shows its candidate set before confirmation.

• Shared maintenance execution -- Dashboard, CLI, and MCP use the same cleanup and deduplication planners/executors instead of maintaining separate mutation logic.

Changed

• Fail-closed maintenance actions -- execution requires a signed, project-bound preview and rejects stale candidate sets with a refreshable conflict instead of applying a partial plan.

• Conservative cleanup noise detection -- normal engineering terms such as compatibility, handoff, sandbox, and version are no longer treated as noise; only explicit fixture markers and tool receipts are opt-in archive candidates.

• Dashboard documentation -- added the canonical Dashboard maintenance, scope, and safety contract to the public documentation map.

[1.6.3] - 2026-08-18

Fixed

• Safe home-directory boundaries -- project binding and SQLite now share the same effective HOME/USERPROFILE resolution on Windows, so an agent launched from a user directory is rejected consistently and cannot create a decoy ~/memorix.db. Contributed by @RaviTharuma in #230, with a maintainer compatibility follow-up for Windows environment resolution.

• Fast first session on large stores -- memorix_session_start uses a bounded newest-first slice while the remaining hydration work continues in the background, keeping first contact responsive on large projects.

[1.6.2] - 2026-08-18

Changed

• Responsive large-store control plane -- HTTP health and MCP initialize no longer wait for full observation or embedding-cache hydration. Hook writes persist directly without loading the full corpus, one-shot CLI paths avoid cached-vector attachment, and ordinary embedding-disabled writes stay off the provider/configuration hot path. Contributed by @RaviTharuma in #214.

• Bounded and crash-safe embedding cache -- JSONL cache loading is bounded by a 256 MiB raw-vector payload budget, legacy JSON migration publishes via temporary-file rename, concurrent loads are single-flighted, and sidecar metadata avoids rescanning a large cache on every save.

Fixed

• SDK lifecycle and persistence parity -- MemoryClient.close() releases its SQLite handle on Windows, reopening the same directory reloads a fresh store snapshot, and SDK clients use the same flat data directory as CLI, hooks, and HTTP MCP instead of the retired per-project directory layout.

• Correct background entrypoint -- memorix background start launches the built HTTP CLI entrypoint rather than the stdio library entry.

[1.6.1] - 2026-08-17

Fixed

• Bounded hook lifecycle on every desktop platform -- memorix hook now reads stdin with an idle bound, performs embedding writes in the deferred lane, and has a hard process deadline. The hook-specific V8 heap is 512 MiB instead of the control plane's 4 GiB. This prevents hung hook processes and associated CPU or memory pressure on macOS, Windows, and Linux. The Windows CLI heap restart also hides its child console window.

• Recoverable HTTP MCP sessions -- HTTP MCP sessions now default to a 12-hour idle window. An expired or unknown session receives a fast 404 with a reinitialize hint instead of looking like a silent client hang. Set MEMORIX_SESSION_TIMEOUT_MS=0 only when the host owns lifecycle cleanup.

• Truthful embedding diagnostics -- the HTTP control plane exposes its real runtime embedding state. memorix doctor reads that state instead of reporting a fresh local probe as ready while the running service is degraded by authorization, billing, timeout, or upstream failure.

• Accurate doctor JSON health -- a live PID with an unreachable HTTP control plane now reports healthy: false rather than a false positive.

Changed

• macOS release coverage -- normal CI now runs the full build and test suite on macos-latest in addition to Ubuntu and Windows.

[1.6.0] - 2026-08-17

Added

• Bounded context receipts -- memorix context|resume --brief-json, memorix codegraph context-pack --brief-json, and MCP project-context format: "receipt" return only the agent-ready brief, selection receipt, CodeGraph quality, and optional Agent loadout. The detailed --json and MCP format: "json" forms remain available for diagnostics.

• Task-scoped Agent loadouts -- --agent <target> selects only compatible project workflows and reports their required context and workflow tools. This is guidance, not a replacement for the host agent's authorization.

• Explicit external CodeGraph lifecycle -- memorix codegraph init|sync manages an already installed local CodeGraph index. It never invokes CodeGraph's agent installer or writes Agent configuration.

Changed

• Honest CodeGraph status -- Status now distinguishes the persistent Lite Code State index from the provider actually usable for the current task. A stale external graph falls back to Lite instead of being presented as fresh.

• Bounded fallback guidance -- Newly generated Agent rules teach --brief-json for the one CLI fallback; --json is reserved for deliberate detailed diagnostics.

[1.5.1] - 2026-08-16

Fixed

• Bounded formation cancellation -- Formation deadlines now abort the LLM and vector-search work they started, including response-body reads and retry waits. A provider that ignores cancellation still cannot hold the caller past its deadline.

• Embedding error classification -- Only an explicit HTTP 400 batch-shape error may trigger batch splitting. Authentication, billing, quota, and upstream errors fail once instead of multiplying latency through recursive retries.

[1.5.0] - 2026-08-14

Added

• Stability stress suite -- Five deterministic, offline stress exams now run in the normal test suite: a 2,000-observation corpus (planted needles, ranked determinism, project isolation), 60-way parallel writes plus 20-way same-topic upserts, session churn across projects with parallel starts, 300-record long-term maintenance exactness, and repeated CLI invocations. Live embedding stress exams are env-gated (MEMORIX_RUN_LIVE_EMBEDDING_TESTS=1) and never run in CI.

• Provider-aware embedding default -- When the embedding base URL points at OpenRouter and no model is configured, the default is now qwen/qwen3-embedding-8b (4096 dimensions) instead of the OpenAI-only default. An explicitly configured model always wins.

Changed

• Verified OpenRouter embedding lane -- The api embedding provider was exercised live against qwen/qwen3-embedding-8b on OpenRouter: a 100-text batch, cache round-trip, and provider-singleton stability all pass. Docs cover the env-var form and borrowing OPENROUTER_API_KEY.

• Multimodal vision option documented -- Image analysis can run on the LLM lane through an OpenAI-compatible vision endpoint, including OpenRouter models such as qwen/qwen3-vl-8b-instruct.

[1.4.6] - 2026-08-14

Added

• Memory-native brief block -- Every task brief opens with a "You and this workspace" block: the user profile (personal user-profile observations), the latest ended session summary, and recent durable long-term memories, all inside the existing token budget. Older callers and fixtures without the block are unaffected.

• Long-term memory self-maintenance (rule leg, no LLM required) -- Explicit long-term requests (MCP longTerm: true, CLI add/promote) auto-qualify on the spot instead of waiting for a manual review; hook-captured and Git-derived candidates keep the review path. Stale qualified/approved records auto-archive after 30 days without activity, and a newer record auto-supersedes same-title records in the same scope and kind. Maintenance runs in the existing durable lane at session start, after each explicit promotion, then on a daily heartbeat. Approval stays an explicit operator action.

• memorix purge -- Retires the current project's memories from retrieval (--all for every project, --yes required outside a terminal). Observations and long-term records are archived, never hard-deleted, so the audit trail survives.

• Eval coverage -- New deterministic, offline exams pin the always-on brief, long-term maintenance, profile parity, purge safety, git ingest policy, media schema honesty, and server config defaults.

Changed

• Installed tool profile matches taught tools -- memorix setup and every plugin template now run memorix serve --mode lite, so the tools taught in generated guidance (sessions, timeline, reasoning, retention) actually exist. Repo guidance marks the full-only tools (promote, rules_sync, workspace_sync) with MEMORIX_MODE=full.

• Honest media schema -- With the generation/transcription gates closed, memorix_media no longer advertises generate-image, generate-video, or derive-audio (or their parameters); agents see only what the operator enabled.

• Config that does something -- [server].port and [server].dashboardPort are now real startup defaults for serve-http and dashboard; git.ingest_on_commit actually gates post-commit ingestion and git.max_diff_size caps captured diff content. The never-consumed mcpServers key was removed from memorix.yml (unknown keys are ignored), and [team]/[hooks] are documented as parsed-but-reserved.

Fixed

• Guidance hygiene -- Generated agent guidance now teaches store-the-underivable, dual-channel feedback, verify-before-use, the ignore contract, user-profile capture, and session-end summaries.

• Session search dedup -- memorix_search demotes rows already shown in the same session (duplicate rate 25% → 19% in the replay exam, hits and token budget unchanged).

• Dead code removal -- Removed the superseded slash-REPL, the legacy in-memory team layer, retired TUI components, and duplicate commands (about 4,600 net lines), each verified unreferenced before deletion.

[1.4.5] - 2026-08-14

Added

• DeepSeek Harness integration -- memorix setup --agent dsh installs Memorix into DeepSeek's official harness: a @deepseek-ai/dsh-mcp-client row in $DSH_HOME/cordis.patch.yml (tools appear as mcp__memorix__*), AGENTS.md guidance at the harness and project levels, and official skills under $DSH_HOME/skills or .dsh/skills. Existing user patch rows are preserved and installs are idempotent. DSH has no hook surface, so none is written. doctor, repair, uninstall, and setup --list cover the new target, and MCP checks for user-level-only adapters now run under an explicit global scope (fixing the same latent skip for Trae). Native validation composed the generated patch with the real @deepseek-ai/[email protected] launcher.

[1.4.4] - 2026-08-14

Added

• MiniMax image-to-image -- memorix media generate image --image <path> sends a reference image through MiniMax's documented subject_reference API for the image-01/image-01-live models. The gated memorix_media MCP tool accepts an optional base64 reference image with the same 20 MiB visual limit and MIME validation. The existing MCP generation billing gate is unchanged. Contributed by @octo-patch in #178.

Fixed

• CLI argument crash on long text -- memorix memory store --text could fail with Error: named?.trim is not a function when a repeated flag or a Windows PowerShell 5.1 argument split delivered the value as an array. CLI argument coercion now joins repeated free-text arguments and takes the last value for scalar and enum flags, with regression tests. Reported by @skywalker-35 in #194.

[1.4.3] - 2026-08-10

Added

• Controlled document and audio derivations -- Explicit CLI jobs can now derive searchable text from managed PDFs and from supported audio transcription providers. The original asset remains local and provenance, provider credentials, retries, cancellation, and provider cost reporting are bounded rather than inferred.

• Evidence-bounded graph context -- Retrieval can add direct, one-hop entity and commit evidence to an existing workset, while using the same freshness, trust, visibility, and token-budget governor as normal project knowledge. It does not perform unbounded graph traversal.

• CodeBuddy integration -- memorix setup --agent codebuddy installs an official-format local marketplace plugin with MCP, skills, and lifecycle hooks, without changing CodeBuddy model, permission, or user settings.

Fixed

• YAML dependency security -- Replaced the unmaintained gray-matter / js-yaml 3 parsing chain with the maintained yaml parser for Memorix configuration, rules, workflows, and knowledge-page front matter. Memcode's HTTP client and the MCP SDK's patched transitive network dependencies are pinned to their fixed releases.

• Optional local Transformers install -- Memorix no longer installs @huggingface/transformers by default because its current upstream release still brings a high-severity sharp/libvips advisory. The local Transformers embedding provider remains available through an explicit user install.

• Star History release verification -- Generated chart pull requests now explicitly dispatch the normal CI workflow. This avoids relying on the GitHub GITHUB_TOKEN behavior that otherwise suppresses a follow-up pull_request workflow.

• Plugin release metadata -- CodeBuddy now participates in the same version-manifest synchronization as the other versioned integration plugins.

[1.4.2] - 2026-08-06

Added

• Evidence-governed memory -- A deterministic governor now decides whether project knowledge should be included, compacted, deferred, or excluded based on source backing, freshness, conflicts, quality, scope, and token budget.

• Outcome-aware memory -- Workflow verification results are recorded as append-only outcome signals so failed evidence is less likely to be reused as trusted project knowledge.

• CodeGraph evolution views -- Lite CodeGraph now keeps hash-only snapshot manifests, reports exact file-level diffs, and exposes bounded one-hop impact slices through the CLI. Continuation worksets include code evolution and stale-memory binding counts.

Fixed

• Star History automation -- The chart workflow now updates a dedicated branch and opens or refreshes a pull request instead of trying to push directly to protected main, which caused every scheduled refresh to fail.

[1.4.1] - 2026-08-03

Added

• MiniMax M3 video input -- The bundled @memorix/ai provider layer now advertises and routes MiniMax M3 video content through MiniMax's Anthropic-compatible API. Base64 video is covered by focused request-shape tests, and image-only models still receive a clear text fallback instead of an invalid video block. Contributed by @octo-patch in #153.

Changed

• Maintained product roadmap -- Replaced the stale version-by-version roadmap with current product boundaries, operating constraints, and public directions. Release facts now live in this changelog and active maintainer work lives in ACTIVE_WORK.md. Contributed by @FBISiri in #140.

Fixed

• Deterministic provider tests -- @memorix/ai unit tests no longer discover local API keys or Pi OAuth credentials and unexpectedly call paid, changing provider endpoints. Routine tests are offline by default; deliberate live validation requires MEMORIX_AI_LIVE_TESTS=1 plus the required credentials.

• Stable model metadata assertion -- Removed a test-only assumption that OpenRouter will permanently list moonshotai/kimi-k2.6:free; the bundled catalog snapshot no longer contains that route, while coverage remains for the supported Kimi K2.6 compatibility metadata.

[1.4.0] - 2026-08-02

Changed

• Stable multimodal release -- Promoted the controlled media lifecycle, explicit MiniMax image/video generation, and honest multimodal embedding foundation introduced in 1.3.3 to the stable line after the 1.3.4 Windows, MCP binding, cancellation, cleanup, configuration-migration, and fresh-package hardening work.

Verified

• Release evidence -- Full local regression, fresh npm-package MCP smoke, GitHub Actions on Windows and Ubuntu, Docker control-plane startup, native SQLite compatibility, npm provenance publication, and official MCP Registry publication all passed before this release.

• Release boundary -- “Stable” means no known release blocker after those gates. It does not claim that any software is mathematically free of future defects.

[1.3.4] - 2026-08-02

Fixed

• Fail-closed project binding -- Stdio and HTTP MCP servers no longer restore or keep writing a remembered project root. A launch without a reliable workspace now waits for an explicit root, MCP Roots, or memorix_session_start, preventing cross-project memory access.

• Portable MCP Roots -- File URI roots now remain POSIX paths on Linux/macOS and convert correctly on Windows, including drive-letter paths and UNC roots.

• Atomic media attachments -- A media retrieval observation and its asset link now commit in one SQLite transaction. If linking fails, no dangling observation survives. Completed video downloads also persist their local asset before a retryable attachment step.

• Authoritative video cancellation -- A cancellation that lands while a video worker is downloading or finishing can no longer be overwritten back to completed; MCP can cancel an existing job even when new MCP media generation is disabled.

• Recoverable media cleanup -- Files Windows cannot immediately release are visibly staged for retry, and media cleanup reports reclaimed or still-pending staged bytes instead of silently implying deletion succeeded.

• Non-destructive Codex migration -- Legacy Codex hook cleanup now preserves customized hook files instead of deleting a file merely because its commands begin with memorix hook.

• One stdio runtime -- Direct node dist/index.js invocation now delegates to memorix serve; importing the package no longer starts an MCP server as a side effect.

Changed

• Bounded default MCP surface -- The micro profile exposes nine tools: agent-ready context, retrieval, one explicit session recovery path, and one controlled media operation. Advanced tools remain opt-in through higher profiles.

[1.3.3] - 2026-08-02

Added

• Controlled media assets -- Added a local content-addressed asset library for explicit images, audio, video, and PDFs. Assets live outside the Git worktree, keep validated MIME/hash/size provenance, support explicit attachment to normal memory, and have link-aware quota cleanup.

• MiniMax media generation -- Added explicit MiniMax image generation and durable MiniMax-H3 video jobs. Video tasks survive process restarts, expose status/cancellation, and import completed output only after bounded safe download validation.

• Honest multimodal embedding foundation -- Added typed media embedding inputs, profile/modality compatibility checks, and separate media-vector storage. Text-only embedding providers now return a capability fallback instead of inventing a visual vector.

• One MCP media operation -- Added memorix_media for agent-facing import, attachment, inspection, and job status while retaining CLI as the complete management surface.

Changed

• Legacy image ingestion joins the asset lifecycle -- memorix ingest image and the full-profile compatibility tool now keep a verified controlled image asset alongside the text retrieval projection instead of discarding the binary.

• MCP binding boundary -- Project binding is now a transport-neutral request context. Legacy HTTP session IDs only route transport requests; explicit project roots cannot be overwritten by Roots discovery, including a workspace that contains a nested Git project.

Fixed

• Atomic media removal -- Asset removal stages a file in a controlled recovery directory, then atomically detaches links, removes derivations/vectors, and marks the asset deleted. A database failure restores the file instead of silently losing it.

• No ambient paid generation -- MCP image/video generation is disabled until MEMORIX_MCP_MEDIA_GENERATION=1 is explicitly set. CLI generation remains deliberate and auditable.

• Bounded image analysis -- Legacy base64 image ingestion validates canonical input before decoding, caps visual analysis at 20 MiB, and sanitizes provider-derived descriptions and diagnostics before persistence.

• Test data isolation -- Tool-profile integration tests now always use temporary Memorix data directories instead of a developer's real local memory database.

• Smaller install artifact -- Excluded runtime-unneeded distribution sourcemaps from the npm package while retaining compiled code, source, documentation, and memcode runtime assets.

[1.3.2] - 2026-08-02

Fixed

• Windows background lifecycle -- Background Control Plane startup now uses a hidden, independent Windows launcher that returns the real Node PID immediately, writes managed state before readiness checks, and avoids both console flashes and host-reclaimed child processes.

• Safe stale-process handling -- Background start, stop, and ensure no longer kill a live PID merely because old local state is unhealthy. A forced stop happens only after the health endpoint proves the PID belongs to the registered Memorix service.

• Truthful setup preview -- memorix setup --dry-run, including --global, now performs no file, plugin, hook, MCP, or agent-setting changes. This is covered for Codex user-level paths.

• CLI failure boundaries -- background, serve-http, and dashboard now reject malformed or out-of-range ports before starting a service; background log line limits are bounded, and these expected input errors are concise rather than internal stack traces.

• Recoverable orchestration work -- Failed or timed-out orchestration worktrees retain uncommitted work for inspection rather than being force-removed; Git helper calls use argument arrays and bounded inputs.

• Dependency hardening -- Refreshed the MCP SDK and transitive package overrides; production dependency audit is clean.

[1.3.1] - 2026-07-31

Fixed

• Windows CLI window flicker -- Memorix now hides Windows child-process windows across normal CLI work: startup restart, Git/project probes, hook runners, setup checks, background controls, orchestration, and memcode utility commands no longer flash a console for each invocation.

• Recoverable vector backfill -- One-shot vector backfill no longer launches as a detached Windows process. The durable queue still preserves and retries the work, while unref() lets the foreground CLI return without a visible extra console.

[1.3.0] - 2026-07-30

Added

• Curated long-term memory -- Added source-bound episodic, semantic, and procedural memory records with explicit candidate -> qualified -> approved -> archived/superseded lifecycle states, evidence references, audit events, and task-aware delivery.

• Deliberate local portability -- A user may explicitly create a user + portable memory for reuse across projects on the same local installation. Project and team facts remain project-bound, and source-derived evidence can never be made portable.

• Long-term memory CLI -- Added memorix memory long-term list|show|add|promote|qualify|approve|archive|supersede for reviewable lifecycle management without requiring an MCP connection.

Changed

• Bounded Worksets -- Task context can now select up to three relevant qualified or approved long-term records, with source and lifecycle status preserved in the context receipt. Candidates, archived records, and superseded records are never injected.

• Low-noise agent guidance -- Generated rules and shipped plugin skills tell agents to create long-term candidates only for stable facts, reusable procedures, or completed episodes; approval remains an explicit operator action.

• Bounded semantic recall -- Keyword matching remains the deterministic first path. A configured embedding provider gets one 1.8-second, no-retry fallback only when no reviewed long-term item matched, so a slow remote service cannot turn context delivery into a long foreground wait.

Fixed

• Long-term CLI discoverability -- memorix memory help now documents the existing --title, --tags, and --applicability fields for curated long-term records.

[1.2.10] - 2026-07-28

Added

• Official MCP Registry publication path -- Ships the checked server.json manifest and a GitHub OIDC publication step that runs after npm succeeds. This release is the first package version eligible for official Registry registration.

Fixed

• Codex plugin ownership -- memorix setup --agent codex --global now uses Codex's personal marketplace plugin path and never creates project-local .codex configuration as a fallback. Custom user MCP configuration stays untouched; legacy Memorix files are migrated only after Codex confirms the plugin is enabled.

• Glama-compatible Docker build -- Docker now installs every workspace dependency set before compiling. CI builds and starts the HTTP control-plane image, so a future container regression is caught on the pull request instead of by an external registry scan.

[1.2.9] - 2026-07-28

Fixed

• One retention truth across the product -- Dashboard, HTTP control-plane, and background cleanup now use the same canonical retention projection. Permanent, active, stale, and archive-ready memory states no longer disagree across surfaces.

• More useful session continuation -- Resume and handoff-shaped requests prioritize the prior task, what changed, and remaining work without treating ordinary verification or compatibility language as disposable noise. Delivery stays bounded rather than replaying a transcript.

• Responsive CLI writes with remote embeddings -- memorix memory store now persists the memory and queues durable vector work before starting a detached backfill worker. A slow embedding API no longer keeps a one-shot CLI command open; lexical recall remains available while vector search catches up.

[1.2.8] - 2026-07-27

Fixed

• Recoverable vector backfill -- Temporary embedding or index failures no longer exhaust a shared background job. Memorix keeps one diagnosable retry with bounded backoff, batches provider requests, and lets a later healthy MCP session resume the same recovery work instead of accumulating permanent failures.

• Accurate vector status outside MCP -- The standalone dashboard and HTTP control plane no longer present an unhydrated in-memory index as a healthy 0 / 0 result. They now explicitly say that vector status belongs to the active MCP session when they cannot observe it.

• Codex plugin ownership -- Agent Doctor recognizes the enabled Codex plugin as the owner of memorix serve, reports first-use hook approval as a host consent step instead of a broken install, and avoids suggesting a redundant config repair.

• Safe Codex migration -- memorix setup --agent codex --global removes only the known old source-path Memorix MCP entry after the official plugin installation succeeds. Custom user-managed MCP entries remain untouched.

[1.2.7] - 2026-07-27

Added

• Native compaction continuity -- Memorix records host-native compaction lifecycle checkpoints separately from durable memory. Pi and Oh-my-Pi preserve the native summary fields their extension API exposes; Codex and Claude Code keep an honest lifecycle marker when their hook payload exposes no summary.

• One-time host recovery -- Codex receives one bounded checkpoint through its documented SessionStart compact path. Claude Code receives the same bounded checkpoint through its documented next UserPromptSubmit context path. Neither path replays a transcript or repeats delivery.

• Claude handoff routing -- Claude Code now receives a compact official UserPromptSubmit hint for handoff and continuation requests, directing one Memory Autopilot call before broad file or Git exploration without injecting a broad brief into unrelated prompts.

• Cross-agent continuation evidence -- Explicit memorix resume / task continuation Worksets can include one recent, source-labelled compact checkpoint from the same Git project. It remains historical lifecycle evidence, not promoted knowledge.

• Checkpoint CLI and advanced MCP inspection -- Added memorix checkpoint list|show|context|archive; the matching memorix_compaction_checkpoint MCP action is available only in the full profile.

Fixed

• Repeated compact lifecycle safety -- Retried pre-compact and post-compact hook delivery is idempotent. Multiple real compactions in one long host session retain separate checkpoints when the host exposes a new preflight marker or native compaction event ID.

[1.2.6] - 2026-07-27

Added

• Truthful Memcode command discovery -- The interactive command list and editor autocomplete now derive from the same executable Pi-compatible command registry. Memorix-native /memory actions, including nested completion after /memory , are available without advertising dead commands.

• Deterministic memory resolution -- /memory delete <id> now resolves the selected project-memory entry instead of only showing a candidate list.

• Pi-compatible shell session context -- Model-initiated Memcode bash calls now receive current PI_SESSION_ID, PI_SESSION_FILE, PI_PROVIDER, PI_MODEL, and PI_REASONING_LEVEL metadata. Extensions can explicitly disable this exposure.

• Maintainer Pi-upstream audit -- Added a pinned upstream manifest, npm run audit:pi-upstream, and a documented review policy. The read-only audit detects GitHub compare truncation and falls back to a full official Git-tree comparison.

Fixed

• No misleading terminal actions -- Removed stale, non-dispatched TUI entries and unreachable /git actions, including the former implicit stage-all commit path.

• Memcode release identity -- Version-check and user-agent internals now identify the published Memorix/Memcode package rather than incorrectly calling it a Pi release; bundled memcode --version now reports the installed Memorix release instead of an internal workspace version.

• Task-flag parity -- memorix context --task "..." now works without a redundant positional task argument.

• Plugin release metadata -- All shipped versioned plugin manifests now match the published Memorix release, including Claude Code, Codex, Copilot, Gemini CLI, OpenClaw, Oh-my-Pi, and Pi.

• Complete Node SQLite fallback -- The supported Node built-in SQLite fallback now matches the transaction variants and permissive named-parameter behavior Memorix uses with better-sqlite3. Team task claims, Knowledge Workspace review, and other row-object writes no longer fail when the optional native binding is unavailable.

[1.2.5] - 2026-07-27

Added

• Explicit retrieval profiles -- Search now accepts fast, balanced (default), or thorough through the CLI, MCP, and SDK. Fast is fully local, balanced keeps optional LLM work out of everyday retrieval, and thorough explicitly opts into configured LLM query refinement.

• Reproducible retrieval evidence -- Added npm run benchmark:retrieval for a clearly-labelled hot in-process SDK lexical benchmark with p50/p95/p99 output. It does not present CLI startup, MCP transport, remote embedding, or LLM latency as the same metric.

Changed

• Lighter read-only CLI path -- Memory search, detail, timeline, recent, graph context, and help no longer create session bookkeeping or maintenance-target metadata. Identity and visibility checks remain unchanged.

Fixed

• Completed searches no longer retain timeout watchdogs -- Embedding, rerank, MCP search, and maintenance timeout paths share a helper that clears the timer when work settles. This removes the avoidable CLI exit delay caused by an already-completed optional rerank.

• OpenRouter lane isolation -- An embedding-only OPENROUTER_API_KEY no longer silently enables the memory LLM lane. It remains valid for a memory LLM only when that lane explicitly selects an OpenRouter provider or base URL.

[1.2.4] - 2026-07-26

Added

• One-call continuation for CLI-only agents -- Added memorix resume "<task>", a direct terminal entry that returns the same bounded Memory Autopilot Workset as Project Context without requiring MCP discovery or command probing.

Changed

• Continuation is a delivery choice, not a separate memory silo -- Project Context now recognizes continuation language independently of its task lens. It adds only the latest meaningful session summary and up to three readable durable anchors, with source records in the existing context receipt. Ordinary new tasks remain free of old-session dumps.

• Stable agent fallback guidance -- Generated skills and rules now direct an agent with unavailable MCP to make exactly one task-aware CLI call: memorix resume for prior work or memorix context for a new task. Continuation requests must take that brief before local file/Git archaeology, and an absent .memorix directory is not treated as proof that no durable memory exists. This prevents broad command enumeration, repeated search loops, and false cold-start conclusions.

Fixed

• Claude continuation delivery -- Claude Code now receives an explicit continuation's bounded prior-work Workset through its official UserPromptSubmit.additionalContext response, rather than relying on an advisory rule or a SessionStart message the host does not place in model context. memory.inject = "silent" still disables automatic injection.

• Fresh-install SQLite resilience -- When the optional better-sqlite3 native binary is missing on a supported Node 22 runtime, Memorix now opens the same local memorix.db through Node's built-in SQLite instead of losing session and durable-memory reads across CLI or hook processes. No data migration or configuration change is required.

• Standard TOML literal strings -- Project and global configuration now accepts single-quoted TOML strings, including # characters inside the literal value.

• MCP continuation summary parity -- memorix_project_context now includes the same bounded prior-session and durable-memory evidence in format: "summary" as it does in the default prompt format, so agents that request a compact summary do not lose a valid handoff.

• Actionable continuation anchors -- Bounded continuation text no longer cuts a technical flag or symbol halfway through. The shared guidance also avoids redundant Context Pack/search/detail calls after a complete Autopilot brief, keeping the default agent path compact.

• Autopilot duplicate retrieval -- A completed MCP Project Context now holds a short-lived delivery boundary. Follow-up search/detail calls do not re-send memories already represented in that brief unless the user explicitly asks to inspect the underlying record.

• Read-only memory mutation -- A task that asks for read-only work or says not to modify files now rejects automatic memorix_store writes. An agent must have an explicit user request before it can override that boundary.

• Prior-work privacy parity -- Continuation retrieval applies the same project, team, and personal-memory visibility reader as session context, including canonical project aliases. A different identity cannot receive another agent's personal durable memory through the new delivery path.

• Rule-surface drift -- The shared automatic-memory rule file now carries the same one-call CLI fallback contract as setup-generated agent guidance.

[1.2.3] - 2026-07-25

Fixed

• Session context now respects observation visibility -- Automatic session handoff and manual session-context reads in the CLI, MCP server, and Workbench apply the caller's project, team, and personal-memory reader. A new identity cannot receive another agent's personal observations through session startup.

• Identity before handoff -- CLI and MCP coordination sessions establish their explicit identity before assembling prior context, so an owner's own personal records remain available while a newly joined agent fails closed.

• Graph-side disclosure guard -- Agent-facing session handoff no longer derives graph-neighbor hints from graph relations that do not yet carry observation visibility metadata.

[1.2.2] - 2026-07-25

Added

• Direct CLI control plane -- Added a consistent terminal surface for memory, context, Code State, knowledge, coordination, audit, and transfer work. --cwd selects the Git project from any shell, memorix workbench explicitly opens the interactive terminal UI, and every action group now has task-oriented help.

• Explicit local CLI identity -- Added memorix identity status|join|use|clear. A user can deliberately activate one project coordination identity for personal/team memory and task, message, lock, handoff, and poll commands without needing an MCP connection.

• Automation-friendly transfer -- Memory exports can write directly to a file, and imports accept --file or --stdin as well as existing inline JSON.

Changed

• One visibility reader across terminal surfaces -- CLI commands, Workbench search, recents, health, graph, knowledge, and chat now resolve the same project/actor reader. An unbound terminal remains project-scoped by default.

• CLI ergonomics -- Root search, remember, and recent aliases now use the canonical memory commands; kebab-case flags are accepted alongside the existing camelCase forms.

Fixed

• Private evidence cannot become public indirectly -- Personal and team observations are rejected when promoting shared skills or generating project skills.

• TUI visibility mismatch -- The interactive terminal UI no longer falls back to an unbound reader after an explicit local identity has been activated.

• Transfer visibility bypass -- CLI and MCP exports now include only observations readable by the current caller, rather than exporting personal/team records from the same project by default.

[1.2.1] - 2026-07-19

Added

• Review-gated Knowledge claims -- Explicit agent observations now become source-backed candidates first. They can be inspected and deliberately approved or rejected through memorix knowledge claims / memorix knowledge review or the advanced memorix_knowledge MCP action before they enter knowledge compilation.

• Versioned Memorix release workflow -- Added docs/knowledge/workflows/memorix-release.md, a canonical release playbook with verification gates and an explicit maintainer approval boundary.

Fixed

• Claude setup respects --noHooks -- Project setup now keeps generated CLAUDE.md guidance without also creating Claude lifecycle-hook configuration when hook capture was explicitly disabled.

• Claude Code manual MCP readiness -- Manual setup examples now include Claude Code's alwaysLoad: true entry and point to Doctor/Repair for detecting and restoring the eager-load contract.

• Workflow import fidelity -- Canonical Windsurf workflows preserve their source ID, title, agents, phases, and verification gates instead of being reduced to a generated workflow:<hash> entry. Release workflows no longer match a non-release task merely because it says “verify” or “test”.

• Graph surface consistency -- Explicit relatedEntities now persist as graph edges, and MCP graph tools, standalone Dashboard, HTTP control plane, and exports share one project-scoping rule.

• Intent-aware task and workflow routing -- A safety constraint such as “do not publish” no longer routes an incident or debugging task into a release lens or release workflow. Explicit release requests still retain release verification while publication is deferred for approval.

• Git fact consistency -- Project Context, Context Pack, and CodeGraph CLI now report Git: unavailable for an invalid or unreadable repository instead of presenting it as a clean worktree.

• Windows verification-gate timeout -- A timed-out orchestration gate now resolves promptly while its shell process tree is terminated in the background, instead of waiting indefinitely for a descendant process to close.

[1.2.0] - 2026-07-18

Added

• Versioned Code State -- Added local code snapshots, source epochs, completeness metadata, and memory-to-code freshness so project memory can be requalified when the checkout changes.

• Claim ledger and reviewable knowledge workspace -- Added source-backed claims with confidence/conflict lifecycle, plus local or explicitly versioned Markdown workspaces that compile proposals before a reviewed page can change.

• Canonical project workflows -- Added import, selection, preview, safe adapter application, and run receipts for project workflows without treating an agent-specific instruction file as the source of truth.

• Bounded task Worksets -- Memory Autopilot and Context Pack now select task-relevant current facts, evidence, knowledge/workflow starts, cautions, and verification instead of adding a generic historical-memory dump.

• Advanced Knowledge Workspace MCP tool -- Added memorix_knowledge as one action-based management surface for workspace, proposal, and workflow operations in the explicit team and full MCP profiles.

• Optional local semantic CodeGraph provider -- A healthy pre-indexed local CodeGraph can add a validated, bounded semantic outline to a task Workset. The built-in Lite provider remains available with explicit capability limits.

Changed

• Durable knowledge lifecycle -- Code State refresh, claim derivation/requalification, knowledge compile/lint, and workflow indexing run through resumable maintenance jobs rather than turning interactive requests into a corpus-wide foreground scan.

• Provider quality is visible -- CodeGraph status, Doctor, and Project Context JSON report whether the task used Lite structural evidence or a validated external semantic outline.

Fixed

• Incomplete scan truthfulness -- Unreadable files and deferred removals now remain visible in snapshot completeness instead of being silently treated as a complete code view.

• External CodeGraph safety boundary -- Memorix only accepts local output for the exact healthy project root, rejects stale/malformed/oversized/path-escaping data, runs without a shell, and never persists raw external source output.

• Node 26 SQLite runtime -- Upgraded the optional better-sqlite3 path to a release that supports Node 26, and added a Node 26 CI smoke that opens an in-memory database. Fixes #130; reported by @RaviTharuma.

[1.1.13] - 2026-07-17

Added

• Codex installation proof -- memorix doctor agents --agent codex --scope global now checks the local plugin bundle, Personal marketplace entry, five declared and trusted lifecycle hooks, and the installed/enabled state reported by codex plugin list.

Fixed

• Codex plugin version drift -- Global setup now stamps the copied Codex plugin manifest with the installed Memorix version, so Codex no longer reports the old template version after an upgrade. The source template is covered by a release regression test as well.

• Workspace package publish trap -- @memorix/ai, @memorix/agent-core, @memorix/tui, and @memorix/memcode are now explicitly internal workspaces. The root memorix package ships their bundled runtime, and the release workflow publishes only that supported public package.

[1.1.12] - 2026-07-17

Added

• Codex lifecycle hook capture -- The Memorix Codex plugin now bundles the documented SessionStart, UserPromptSubmit, PostToolUse, PreCompact, and Stop hook events, with a Windows-safe commandWindows command override.

Fixed

• Codex hook normalization and injection -- Codex lifecycle payloads now preserve their event names, SessionStart returns Codex-native hookSpecificOutput.additionalContext, and capture-only events stay quiet instead of filling the agent context with save-status messages.

• OpenCode hooks on Windows -- The generated plugin resolves the installed npm memorix.cmd shim while setup runs and uses that stable path inside OpenCode, instead of relying on OpenCode's inherited PATH. Hook delivery stays quiet unless MEMORIX_HOOK_DEBUG=1 is explicitly set. Fixes #125.

• Setup scope isolation -- Project-level setup no longer installs Claude, Codex, or Copilot plugin bundles in the user home. Use --global for user-level plugins, skills, and lifecycle hooks.

• TUI embedding startup -- TUI health, chat, and quick-search paths now use the same lexical-ready startup flow as MCP instead of making a remote dimension probe during the first screen or search.

• Semantic recall after restart -- API-backed embeddings now restore compatible vectors from the local cache during index hydration, without placing remote embedding calls on the MCP startup path. Redundant namespace-scoped cache metadata survives a lost dimensions cache; cache misses enter the existing bounded background recovery lane, and the first successful vector upgrades a cache-less lexical index in the same process. Based on the diagnosis in #126 by @Tom-Ma-Ming.

• Concurrent index startup -- Cache-only startup and ordinary database callers now share one Orama initialization, so a later vector-schema attempt cannot replace an already hydrated lexical index. Resets also discard obsolete in-flight instances.

• Lazy API vector-cache I/O -- A cache-only start with no trusted dimension metadata now skips parsing the potentially large API vector cache entirely; the cache is loaded only after compatible local metadata or a normal embedding initialization establishes its dimensions.

[1.1.11] - 2026-07-16

Added

• Durable runtime maintenance -- Added a SQLite-backed maintenance ledger for vector recovery, retention, consolidation, and Code Memory refresh. Jobs use dedupe keys, leases, heartbeats, retry backoff, resumable cursors, and operator-visible state.

• Isolated maintenance runner -- Retention, consolidation, and Code Memory scans now run through a compiled child-process runner instead of sharing an MCP request/event loop. Vector recovery stays with the owning process because its Orama index is process-local.

• Code Memory scan safeguards -- CodeGraph Lite now skips common dependency/build directories and source files larger than 2 MiB by default. Configure the limit with [codegraph].max_file_bytes or YAML codegraph.maxFileBytes.

• Maintenance diagnostics -- Dashboard APIs and System Status now show queued/running/failed maintenance state without mixing projects.

Changed

• First-turn MCP readiness -- memorix_project_context, memorix_context_pack, memorix_graph_context, and memorix_codegraph_status now read the current project's SQLite state directly before full Orama hydration. Search, writes, and sessions retain their full-runtime boundary.

• Bounded memory lifecycle -- Automatic retention and consolidation process one project page at a time and resume through durable cursors, avoiding corpus-sized work in an interactive request.

• Incremental Code Memory -- Changed files are reparsed, unchanged files keep their metadata, and removed/oversized files are removed from the graph without replacing the entire project graph.

• Mobile dashboard navigation -- The desktop sidebar becomes a compact mobile navigation layout so dashboard content remains readable on narrow screens.

Fixed

• Cross-process refresh model -- Removed the obsolete JSON-file polling watcher. SQLite generation checks are now the sole authoritative mechanism for cross-process observation freshness.

• Maintenance history and diagnostics -- Completed maintenance records expire after a retention window; persisted failure messages redact credential values.

• Maintenance enqueue atomicity -- History pruning now runs inside the enqueue transaction, so a failed enqueue cannot discard completed diagnostic history.

• Dashboard project scope -- Knowledge, graph, retention, observation, and export reads use project-scoped SQLite queries instead of loading and filtering the full flat memory store.

• Behavior configuration alignment -- TOML/YAML memory behavior settings now drive runtime injection, formation, auto-cleanup, and sync advisory behavior; legacy config.json is a fallback rather than a separate source of truth.

• Container runtime privilege -- The official HTTP control-plane image is explicitly verified to run as the non-root node user.

[1.1.10] - 2026-07-13

Fixed

• Large-project Memory Autopilot timeouts -- memorix_project_context now reuses one CodeGraph snapshot per request and performs first-use code-ref backfill in memory with one batched write instead of per-memory database scans and transactions. Prose-only and ambiguous symbol mentions no longer create large volumes of noisy code references, keeping first-use context generation responsive with thousands of active memories.

• Ruby CodeGraph symbol fidelity -- Ruby namespaces and punctuated method names such as Foo::Bar, save!, valid?, and name= now remain intact through Lite indexing and memory-to-code binding.

[1.1.9] - 2026-07-12

Added

• Configurable CodeGraph excludes -- Added [codegraph].exclude_patterns / YAML codegraph.excludePatterns support so CodeGraph, Project Context suggested reads, context packs, diagnostics, and related CLI flows can skip project-specific generated or vendor paths while keeping the built-in defaults.

Fixed

• Config inspection aliases -- memorix config get now accepts TOML-style snake_case dotted keys such as codegraph.exclude_patterns when reading resolved camelCase config values.

• Orama search index consistency -- Search access tracking now preserves internally stored vector-backed documents, public/detail cache results continue to strip embeddings, and hydration reconciles persisted observations by exact composite ID instead of skipping observation hydration when a shared index already contains mini-skills.

[1.1.8] - 2026-07-08

Added

• Agent integration doctor and repair -- Added memorix doctor agents and memorix repair agents to inspect and repair Memorix-owned agent integration files. The doctor flags stale MCP command paths, missing Claude alwaysLoad, missing memorix MCP entries, and outdated Memory Autopilot guidance without printing environment secrets.

Changed

• Memory Autopilot adoption -- Setup-generated agent guidance now makes memorix_project_context the default first step for non-trivial coding work before progress files, dev logs, broad file reads, or git archaeology.

Fixed

• Claude Code local MCP repair -- memorix doctor agents and memorix repair agents now inspect and repair Claude Code 2.x project-private local MCP entries in ~/.claude.json, replacing stale worktree commands with memorix serve and alwaysLoad: true.

• Cleaner CodeGraph Memory briefs -- CodeGraph Lite and Memory Autopilot now ignore .tmp, nested .worktrees, and .claude/worktrees directories, so suggested reads point at the real project instead of local caches or agent scratch worktrees.

• All-scope doctor noise -- Agent doctor now treats one healthy local, project, or global scope as sufficient in all-scope mode, while still flagging genuinely stale or repairable configs.

[1.1.7] - 2026-07-07

Added

• Task-lensed Memory Autopilot -- memorix context --task ... and MCP memorix_project_context now infer a task lens (bugfix, feature, release, onboarding, refactor, docs, test, or general) and shape Start here files, reliable memories, cautions, and verification hints for that task.

Changed

• Brief noise control -- Release/onboarding/docs-shaped briefs now keep unrelated stale or suspect memory details out of the main prompt while preserving counts and warnings, so agents see less old-context noise.

• Agent guidance -- Generated rules and official skills now tell agents to pass the user's actual task into memorix_project_context or the CLI fallback, making natural-language continuation the default product path.

[1.1.6] - 2026-07-06

Changed

• Release-safe model catalogs -- npm run build and prepublishOnly no longer refresh live model catalogs. Model catalog refresh is now an explicit maintenance action via npm run update-models, so publishing uses the generated catalog already committed to the repository.

Fixed

• Live catalog shrink guard -- Model catalog generators now refuse to overwrite the checked-in catalog when a live API returns an unexpectedly small result set, unless MEMORIX_ALLOW_MODEL_CATALOG_SHRINK=1 is set intentionally.

• Project-scoped diagnostics -- memorix doctor now reports observation totals for the current project instead of mixing in unrelated flat-store project records.

• Claude Code handoff guidance -- Project and plugin instructions now point fresh Claude Code sessions at memorix_project_context, prefer official MCP tool discovery before shell fallback, set alwaysLoad: true for the small Claude plugin MCP surface, and use the repository-wide dev log instead of stale package-specific memcode notes.

• Setup-installed rules refresh -- Re-running setup now refreshes the Memorix-owned block in shared CLAUDE.md / AGENTS.md / GEMINI.md files while preserving user-authored content, so upgraded guidance actually reaches existing installs.

[1.1.5] - 2026-07-02

Added

• Memory Autopilot brief -- memorix context and memorix_project_context now present a compact agent-ready brief with Start here files, reliable code-bound memories, stale/suspect cautions, verification hints, and usage guidance.

• Unbound memory tier -- memorix_context_pack now includes task-relevant memories that do not yet have current code refs as lower-trust unbound context instead of silently dropping them.

Changed

• Context ranking policy -- Context packs now separate reliable current memories from stale warnings and unbound leads, so agents can use memory as an action map rather than a raw text dump.

• Agent guidance loop -- Generated rules and official Memorix skills now teach agents to get project context, inspect suggested files, store durable outcomes, and resolve stale/completed memories.

Fixed

• Agent tool allowlist auditability -- Agent-core now has explicit turn allowlist lookup and regression coverage proving a tool absent from the active turn context cannot execute.

• Session selector search safety -- memcode session search no longer executes user-provided re: input as a regular expression; regex-shaped input is treated as literal search text.

• Autopilot current-truth grounding -- memorix context and memorix_project_context now put live package, changelog, Git branch/commit, and stale progress-note warnings before memory hints so agents do not mistake old dev-log files for the current project state.

• Default install audit surface -- The published package no longer installs fastembed by default because its current releases still pull vulnerable tar versions for downstream consumers. The FastEmbed provider remains available when users install fastembed explicitly.

• YAML parser advisory -- Updated the pinned js-yaml 3.x line to the patched 3.15 release while keeping compatibility with existing gray-matter workflow parsing.

[1.1.4] - 2026-07-01

Changed

• Contributor attribution -- Added post-release credit metadata for the 1.1.3 fixes accepted from #105 and #98 so GitHub can associate the accepted work with the contributing accounts.

Fixed

• Stdio MCP startup latency -- memorix serve now registers MCP tools before loading the full project memory runtime, then warms the runtime in the background. This keeps initialize / tools/list fast for IDEs and registries even when the local memory corpus is large, while real tool calls still wait for project runtime readiness before executing.

[1.1.3] - 2026-06-29

Added

• CodeGraph Memory MVP -- Added a SQLite-backed code structure plane with project-scoped files, symbols, import edges, and observation-to-code references. This lets Memorix store structured code facts beside text observations instead of treating memory as text-only context.

• CodeGraph Lite provider -- Added a built-in TypeScript/JavaScript indexer for file hashes, basic symbol extraction, and import edges, with no new runtime dependency. Richer external CodeGraph providers can be added later against the same store and context-pack contract.

• Freshness-aware context packs -- Added code-ref freshness evaluation and context pack rendering so agents can distinguish current code-bound memories from suspect or stale ones, then inspect suggested files/symbols.

• CodeGraph operator surface -- Added memorix codegraph refresh, memorix codegraph status, and memorix codegraph context-pack, plus MCP tools memorix_codegraph_status and memorix_context_pack.

Changed

• Memory writes attach to code refs -- New observations now best-effort bind to indexed files and symbols after they are stored, without blocking ordinary memory writes if CodeGraph Memory is unavailable.

• Documentation updated for multi-dimensional memory -- API docs and design notes now document CodeGraph Memory as the first context-fabric layer beyond pure text memory.

• OpenRouter embedding key support -- The embedding lane now accepts OPENROUTER_API_KEY when [embedding].base_url points to OpenRouter, while keeping MEMORIX_EMBEDDING_API_KEY as the explicit highest-priority embedding key.

Fixed

• Session project binding -- memorix_session_start({ projectRoot }) no longer hijacks an already-bound parent repo to the first nested git repository under that path. Based on #98 by @Tom-Ma-Ming.

[1.1.2] - 2026-06-27

Fixed

• Removed unowned product-domain defaults -- Removed hardcoded unowned product-domain references from memcode defaults, tests, and generated help text. Version checks now use the official npm registry endpoint, changelog links point to GitHub releases, OpenRouter attribution uses the GitHub repository URL, and /share falls back to the GitHub gist URL unless the user explicitly sets MEMCODE_SHARE_VIEWER_URL.

• Disabled default install report endpoint -- memcode no longer sends install/update pings to any default product-domain endpoint. A report endpoint is used only when MEMCODE_INSTALL_REPORT_URL is explicitly configured by the user.

[1.1.1] - 2026-06-26

Added

• OpenClaw compatible bundle -- Added memorix setup --agent openclaw support for an OpenClaw-compatible bundle with bundled .mcp.json, official skills, and an OpenClaw HOOK.md/handler.ts hook pack, plus workspace MCP sync support for OpenClaw's ~/.openclaw/openclaw.json mcp.servers format.

• Hermes Agent plugin -- Added memorix setup --agent hermes support for a Hermes plugin under ~/.hermes/plugins/memorix, plugin enablement in ~/.hermes/config.yaml, plugin hooks, a slash command, a CLI command, official skills, and mcp_servers MCP config.

• Oh-my-Pi package -- Added memorix setup --agent omp support for an omp.extensions package with extension hook events, a memorix command, official skills, and Oh-my-Pi .omp/mcp.json / ~/.omp/agent/mcp.json MCP config.

Changed

• Gemini CLI and Antigravity positioning -- Documented Google's current Gemini CLI to Antigravity CLI transition while keeping gemini-cli and antigravity as separate Memorix targets because their supported install/config lanes are still distinct.

• OpenClaw/Hermes/Oh-my-Pi setup lanes -- These targets now use their official package/plugin/bundle extension points instead of being treated as MCP-only fallback clients.

Fixed

• MCP config merging -- Setup now deep-merges nested mcp.servers JSON configs and preserves existing Hermes YAML keys when adding the Memorix server.

• Workspace tool schemas -- memorix_workspace_sync now accepts gemini-cli, openclaw, hermes, and omp targets; memorix_rules_sync now exposes the existing Gemini CLI rules adapter.

• Hook fallback guard -- Direct memorix hooks install --agent openclaw|hermes|omp no longer writes fallback hook files and instead points users to the package-owned setup lane.

[1.1.0] - 2026-06-21

Added -- Official Agent Integration Packages

• One-command setup -- Added memorix setup as the main installer for Claude Code, Codex, GitHub Copilot CLI, Cursor, Windsurf, Gemini CLI, OpenCode, Pi, Kiro, Antigravity, Trae, and generic MCP clients.

• Official plugin/package templates -- Added first-party integration packages under plugins/ for Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, Pi, and Cursor package assets. Setup installs each host through its closest native entry point instead of asking users to assemble MCP, rules, skills, and hooks by hand.

• Official Memorix skills -- Added seven practical agent skills: memory search/store, reasoning memory, sessions and handoff, Git Memory, mini-skills, orchestration, and troubleshooting. The skills prefer MCP tools when available and document CLI fallbacks for operator use.

• Integration matrix docs -- Added docs/INTEGRATIONS.md and updated README/setup docs to explain plugin packages, MCP, project guidance, hooks, skills, memcode, and when HTTP MCP is actually useful.

Changed

• CLI as the operator surface -- The memorix CLI now documents and exposes the operator path for setup, memory, reasoning, sessions, Git Memory, transfer, retention, dashboard, orchestration, sync, ingest, skills, and diagnostics. MCP remains the agent-tool entry point.

• Stdio MCP stays the default -- Normal agent setup uses memorix serve. HTTP MCP is documented as an advanced control-plane path for dashboard, Docker, shared endpoint, or multiple clients.

• Agent Team wording tightened -- User-facing docs now frame orchestration as subagent/task coordination through memorix orchestrate, not a separate “agents team” product.

• memcode enters the 1.1 line -- The bundled first-party memagent remains available through memorix, memorix memcode, and memcode, with shared Memorix project memory and updated setup/docs.

Fixed

• Release packaging coverage -- npm package contents now include official plugin templates, skills, hooks, and integration docs so published installs can generate the same setup files tested in the repository.

• Claude marketplace version drift -- The generated Claude local marketplace now uses the current CLI package version instead of a hard-coded release number.

• Session CLI fallback -- memorix session start supports --projectRoot, matching the official sessions/troubleshooting skill guidance for CLI fallback use.

• Windows publish workflow alignment -- GitHub npm publish workflow now uses Node.js 22, matching the package engine floor.

Verification

• Targeted setup, hook, skill, CLI, and operator-surface tests passed during release preparation.

• npm run lint, npm run build, and npm pack inspection were used to verify TypeScript, bundled CLI output, and published package contents.

[1.0.11] - 2026-06-13

Added -- memcode Native Agent Release Candidate

• Bundled memcode entry path -- memorix with no subcommand and memorix memcode now launch the native memcode coding agent path, backed by the Pi-derived agent runtime under the Memorix package namespace.

• Native Memorix runtime awareness -- memcode now exposes first-party runtime status through memorix_status and /memory status, including canonical project identity, shared aliases, project memory counts, source/value distribution, embedding/vector status, search mode, memory LLM lane, retention summary, native hook state, and last injected memory refs.

• Native hook bridge -- memcode feeds its own prompt, tool, and assistant lifecycle events into the Memorix hook pipeline by default instead of behaving like an external MCP-only integration.

• Memory startup and command surface -- restored the normal agent conversation surface with full slash-command discovery, startup diagnostics, startup card, memory commands, and native /memory hooks inspection.

Changed

• Shared project memory remains the default -- memcode uses the same canonical project memory pool as Claude Code, Codex, and other Memorix-connected agents. memcode-specific behavior is represented with metadata rather than a separate memory bucket.

• API key lanes are isolated -- memory LLM, embedding, and agent calls now use separate configuration lanes so users can mix providers without accidental cross-lane key reuse.

• TUI interaction model stabilized -- mouse text selection, wheel scroll, prompt history, activity status, and input rendering were aligned with normal coding-agent TUI expectations.

• Runtime floor aligned with memcode -- the package now declares Node.js >=22.19.0, matching the bundled native coding-agent stack.

Fixed

• Embedding fallback noise -- invalid or unavailable embedding providers now fall back to BM25 with concise diagnostics instead of repeatedly deforming the TUI or spamming runtime output.

• Runtime config loading -- YAML loading works in ESM contexts without relying on an unbound CommonJS require.

• Publish blocker -- removed the stale Pi shrinkwrap path from memcode packaging so @memorix/memcode no longer locks installs to @earendil-works/pi-* packages or calls a missing shrinkwrap generator.

[1.0.10] - 2026-06-08

Added -- Low-Intrusion Handoff Diagnostics

• Privacy-safe handoff receipts (#95) -- Added memorix receipt --json for cross-agent memory debugging. Receipts expose project identity hashes, write/search counts, recent observation ID hashes, and privacy boundaries without raw chat, raw memory text, raw search queries, tool payloads, or local file paths.

• Doctor receipt mode -- Added memorix doctor --receipt so support/debugging can include the same receipt summary alongside normal health diagnostics.

• Cross-agent memory boundary docs -- Documented the operational boundary: MCP connection success does not mean a memory was written, and shared memory means stored memories are searchable across clients in the same Git project, not that chat transcripts are mirrored.

• TUI Agent LLM config scope -- Added agent / MEMORIX_AGENT_LLM_* configuration for TUI/chat agent calls. Memory formation, compression, rerank, and embedding remain on their existing llm / embedding configuration paths. If agent is unset, TUI chat falls back to llm for backward compatibility.

Changed

• Generated agent rules are less intrusive -- Rules now prefer memorix_search when useful and treat memorix_session_start as optional, reserved for handoff, long-running work, session recovery, team coordination, or HTTP project binding.

• Agent Team wording tightened -- Public docs now frame Agent Team as optional autonomous-agent/subagent coordination state rather than an IDE-window chat room. Existing team_* APIs and CLI namespaces are unchanged.

Fixed

• Release publish blocker -- tests/integration/release-blockers.test.ts now reserves an ephemeral local TCP port for the real memorix serve-http smoke test instead of assuming port 19879 is free. This removes a CI-only EADDRINUSE failure that blocked the npm publish workflow.

• Standalone dashboard embedding config (#46 follow-up) -- memorix dashboard now loads project .env / project YAML before dashboard status routes initialize embedding provider state, matching status, doctor, and TUI behavior.

Documentation

• Added a small moving-between-machines note covering memorix transfer export/import without changing the Memorix package name or primary README structure.

[1.0.9] - 2026-05-19

Added -- Knowledge Base / LLM Wiki Foundation

• First-class Knowledge Base layer -- Added a readable project knowledge layer generated from durable memory, git-backed facts, mini-skills, and project evidence. The Knowledge Base is a shared human-and-AI surface, not a replacement for raw observations and not a TUI-only feature.

• Knowledge graph projection -- Added a semantic graph projection over the same eligible knowledge inputs, preserving source refs for navigation and traceability.

• Dashboard Knowledge foundation -- Dashboard surfaces can consume the same generated knowledge contract as the TUI, keeping Knowledge Base and graph semantics aligned.

Added -- TUI Knowledge Workbench

• Tabbed terminal workbench -- Reworked the TUI around Home, Knowledge, Memory, Workbench, and Graph tabs.

• Knowledge browsing -- Added a Knowledge tab for generated wiki sections, item summaries, entities, and source refs.

• Memory browsing -- Added a Memory tab for recent/search/detail flows with ref-focused navigation from Knowledge.

• Workbench session center -- Added explicit session status, explicit bind/end actions, context source summary, and chat. Entering the tab does not auto-start a session.

• Graph text browser -- Added a terminal-native Knowledge Graph tab with summary stats, cluster browsing, node detail, light filtering, and Graph -> Knowledge jumps.

• Cross-surface ref navigation -- Added Knowledge -> Memory, Memory -> Knowledge, and Graph -> Knowledge navigation paths based on stable refs where available.

Changed

• Knowledge architecture clarified -- Documented the memory layer as the primary input to the knowledge layer, while project evidence such as docs, source structure, and git facts act as secondary inputs.

• TUI scope tightened -- The TUI now behaves as a knowledge-native workbench, not a placeholder coding-agent harness. No fake Run Task / Review controls are exposed.

• Session semantics preserved -- session_start remains explicit and lightweight. Agent Team join remains opt-in.

Fixed

• Formation duplicate detection (#94) -- Fixed Formation resolve comparing raw Orama/BM25 ranking scores against normalized similarity thresholds. Noisy log-dump memories with high backend scores no longer discard unrelated valid memories as duplicates.

• TUI chat input lock -- Fixed a regression where after a free-text chat response the bottom input bar became unresponsive because the chat overlay disabled the CommandBar. Chat overlay now keeps input active so users can continue the conversation.

• TUI read-only overlay input lock -- Fixed read-only overlays such as /doctor disabling the CommandBar. Users can now keep typing commands or questions while diagnostics and other non-action views are open.

• TUI shortcut/input conflicts -- Fixed CommandBar conflicts with view shortcuts such as Graph f/k, Memory k, and Knowledge m while still allowing normal first-character input.

• Graph TypeScript build issue -- Fixed a type-only re-export under verbatimModuleSyntax.

Known Limitations

• The Knowledge Base and Knowledge Graph are read-only generated projections in 1.0.9.

• 1.0.9 does not include a desktop app, realtime agent messaging layer, GraphRAG system, graph editing, or full coding-agent harness.

Stats

• Verified TUI subset: 95 passing tests across interaction, workbench, and graph suites.

• Release gates run during cleanup: npm run lint, npm run build, and full npm test.

[1.0.8] - 2026-04-19

Added -- Operator CLI Surface

• Namespaced operator commands -- Added human-oriented CLI namespaces for session, memory, reasoning, retention, formation, audit, transfer, skills, team, task, message, lock, handoff, poll, sync, and ingest so Memorix-native operations no longer require raw MCP tool calls.

• Lightweight session start + explicit collaboration join -- memorix session start now opens project-bound memory sessions without auto-registering a team identity. Operators can opt into collaboration with --joinTeam, or join later via memorix team join, while still getting default role mapping from agentType when they do join.

• CLI-first project ops -- Terminal operators can now inspect memory detail/timeline, store and search reasoning traces, run retention and formation checks, audit attribution, export/import memory snapshots, inspect/generate skills, run explicit rules/workspace sync flows, and ingest images without leaving the shell.

Added -- Programmatic SDK

• memorix/sdk subpath export -- createMemoryClient() factory returns a self-contained MemoryClient with store, search, get, getAll, count, resolve, and close methods. No MCP transport or CLI needed — initialize from a Git project root, read/write observations directly.

• createMemorixServer re-export -- Embed the full MCP server into your own Node.js process and connect it to any transport.

• detectProject re-export -- Standalone Git-based project detection.

• Three subpath exports -- memorix/sdk (runtime API + types), memorix/types (type-only), memorix (MCP stdio entry).

Added -- Official Docker Deployment

• Official HTTP control-plane container path -- Dockerfile now builds a real serve-http runtime image instead of defaulting to stdio MCP.

• Compose example -- added compose.yaml with port 3211, persistent data volume, and /health healthcheck.

• Docker docs -- added docs/DOCKER.md and linked Docker deployment from README / setup docs.

• Runtime truth -- documentation now explicitly states that Docker support is for the HTTP control plane and that project-scoped Git/config behavior requires the container to see the bound repo path.

Visual Semantic Layering (dashboard follow-up)

• Team mental model tightened before release: docs and dashboard copy now say "explicit collaborators" instead of implying a persistent IDE-window roster.

• memorix team status now defaults to active collaborators only; inactive/historical identities remain available with --all for audits without flooding the normal operator view.

• Team page headline is now active-only; recent/historical shown as secondary subtitle only.

• Team agents list gets four filter tabs: Active (default) / Recent / Historical / All. Historical rows no longer flood the view; they sit behind an explicit "Show historical (N)" toggle.

• Each agent row carries an explicit tier badge (Active/Recent/Historical) instead of a single "offline" sea of red.

• Historical agent rows are de-emphasized (reduced opacity, muted colors) so they don't dominate the main view.

• Identity page headline shows real project count + real alias groups only. Temporary/placeholder IDs moved into a collapsed "Historical / temporary project IDs" section.

• Identity "Dirty IDs" card now splits current-project dirty (primary) from historical dirty (secondary) so a single placeholder no longer looks like an active problem.

• Project switcher groups items into Current / Real projects / Temporary / Placeholder, with temporary and placeholder folded behind a "Show temporary" toggle. Each group has a visible count so users can tell a real project list apart from the historical test/smoke scratch pile.

• All new labels wired through the i18n system (English + Simplified Chinese).

Team Page Cleanup (1.0.7 final)

• Team page semantics: clearly presented as a project collaboration space, not an organization backend or staffing admin tool. Scope labels: "Project Collaboration" / "All Projects".

• Explicit collaboration join: memorix_session_start is lightweight by default and no longer auto-registers a team identity. Collaboration is now opt-in via joinTeam: true or team_manage(join), and default role mapping from agentType applies only when joining.

• "Continue This Project" resume area: shows open tasks, available-to-claim tasks, open handoffs, unread messages, active locks, and active agent count at the top of the Team page. Provides a clear "pick up where you left off" entry point.

• Unified statistics: stat cards clearly labeled — Active Agents (with session count + historical total), Locked Files, Tasks (by status), Messages (unread count). No conflicting or ambiguous labels.

Added -- Multi-Agent Orchestrator (Phase 7)

• memorix orchestrate -- Autonomous pipeline: plan → parallel execution → verify gates → fix loops → review → merge. 4 agent adapters (Claude, Codex, Gemini, OpenCode), capability routing with quotas, worktree isolation, evidence collection, stranded task detection, agent fallback on repeated failures.

• CLI flags: --agents claude:2,codex:1, --goal, --parallel, --compile-command, --test-command, --max-fix, --budget, --routing, --memory-capture, --no-evidence, --global-timeout

Added -- SQLite Canonical Store (Phase 2)

• Observations, mini-skills, sessions, archives all in SQLite. Shared DB handle, freshness-safe retrieval, in-place archiving.

Added -- Team Identity (Phase 4)

• Team store with agent registration, heartbeat, task board, handoff artifacts, stale detection. Prompt identity contract, sticky attribution prevention.

Fixed

• HTTP stale-session poisoning (#82) — serve-http no longer has only a hardcoded 30-minute idle GC. Added MEMORIX_SESSION_TIMEOUT_MS so long-running HTTP MCP clients such as Codex can keep sessions alive for longer work blocks without getting stuck on stale Mcp-Session-Id failures.

• TUI responsive layout — sidebar width now scales with terminal width (28% ratio, 26-40 range) instead of fixed 34 columns; content area uses flexGrow to fill remaining space; maximized terminal windows now show expanded UI instead of locked small-window layout

• TUI search score display — replaced absurd raw-percentage display (e.g. "927%") with relative relevance dots (●●●○○) normalized against top result

• TUI /resume numeric index — /resume 2 now selects thread #2 from the list (1-based), in addition to /resume <threadId> by ID

• TUI CJK text wrapping — Ink's wrap="wrap" doesn't understand double-width CJK characters, causing garbled text; replaced with CJK-aware manual line splitting in chat output and ContextRail sizing so mixed CJK content renders predictably

• TUI layout DRY — sidebar/content width calculation extracted to shared computeLayoutWidths() in theme.ts; App.tsx and ChatView.tsx now use the same function, preventing drift

• TUI CommandBar overlay — palette no longer pushes input bar down; rendered as overlay in App.tsx; mouse SGR mode centrally managed to avoid conflicts between palette clicks and chat scrolling

• TUI chat freeze under assistant output — removed assistant Markdown rendering from the workbench chat path and render chat bodies as plain text with CJK-aware wrapping; this avoids heavy reparsing/rerender spikes that could freeze the TUI when the assistant started answering after tool calls

• #4 Parallel multi-agent — fully implemented via orchestrator

• #52 observations.json perf — migrated to SQLite

• #56 LLM rerank timeout — configurable via MEMORIX_RERANK_TIMEOUT_MS

• #75 Cursor stdio binding — deferred-binding mode instead of exit

• Gemini JSON parse — brace-counting extractor handles trailing text

• Evidence for failed tasks — both fix-exhausted and normal failure paths now write evidence

• Agent fallback — failed agents excluded from retry routing

• P1: uninstall hooks — shared context files (AGENTS.md/GEMINI.md) now block-level removal instead of deleting entire file

• P1: budget abort — current settled dispatch now fails task (was left stuck in_progress)

• P1: streaming completion — waits for stdout/stderr close before resolving (was missing final JSONL lines)

• P1: Claude adapter headless hang — added --bare flag to skip hooks/LSP/keychain/interactive ops, plus --mcp-config to restore Memorix MCP tool access in bare mode

• P1: planner JSON materialization — use accumulated text from message stream instead of ring buffer tailOutput (avoids truncation when planner does many tool calls before outputting plan)

• P1: reviewer hanging — --mcp-config enables memorix_handoff / team_task MCP tool calls that reviewer needs for completion path

• P1: structured planner double-create — plannerType: 'plan' tasks now forbidden from calling team_task create (must output JSON only; coordinator materializes). Prevents duplicate engineer/reviewer tasks in goal-mode pipelines

• P2: budget validation — CLI and coordinator reject NaN/negative/zero budget values

• P2: streaming buffer — bounded at 1024 messages with drop-oldest policy

• P2: install audit — recordFile called even when Memorix block already exists (audit self-heal on reinstall)

• P2: install audit — new shared file creation also records audit entry

• P2: uninstall return — rules-only agents (codex/gemini-cli) return true when audit cleanup succeeds

• #80 OpenCode hooks — plugin now uses individual event-name keys (session.created, file.edited, command.executed, etc.) instead of invalid catch-all event handler that OpenCode never called. Replaced fragile cat | memorix hook pipe with Bun.spawn stdin pipe for reliable cross-platform invocation. Added diagnostic logging on failure. Hooks status now distinguishes verified (config-based) from unverified (plugin-based) agents.

• OpenCode plugin v5 — replaced Bun.spawn with child_process.spawnSync for cross-runtime compatibility (OpenCode may fall back to Node.js on Windows where Bun segfaults). Fixed file.edited field extraction (input.path before input.file). Added exit code and stderr diagnostic logging. Added 10s timeout to prevent hangs.

• Copilot Windows runtime — powershell field omitted from hook config when pwsh (PowerShell v7+) is not installed, preventing spawn pwsh.exe ENOENT errors. Copilot falls back to bash field (Git Bash). Install command warns if pwsh is missing on Windows and suggests winget install Microsoft.PowerShell.

• Copilot global hooks — getGlobalConfigPath('copilot') was falling through to the Claude case and returning ~/.claude/settings.json, which is completely wrong. Per official GitHub docs, Copilot only supports project-level hooks at .github/hooks/*.json — there is no global hooks path. Fixed by returning empty string for global path, adding guards in install/uninstall to reject --global for Copilot, and updating hooks status to skip the global check.

• Hook handler diagnostics — runHook() store path and handleSessionStart() now log errors to stderr instead of silently swallowing, making end-to-end pipeline failures visible for debugging.

Added -- 1.0.7 Closeout

• Routing explainability — buildRoutingDecision() / buildIdleReasons() helpers trace adapter selection reasons (default_preference, cli_override, quota_fallback, excluded_failed, last_resort) in pipeline trace, evidence, and summary

• Structured role extraction — extractRole() prioritizes metadata.role over [Role: ...] text parsing; canonical source is structured metadata

• Balanced scheduling — --scheduling balanced policy with round-robin tiebreaker among equally-preferred adapters; best-fit remains default

• Idle agent visibility — pipeline summary includes idleAgents with reasons for non-participation

• Pipeline summary extensions — routingDecisions and idleAgents fields in PipelineSummary for full explainability

• #62+#74 dashboard loadDotenv

• #70 doctor health check

• #69 background start non-TTY hang

• #66 dashboard delete/cleanup/export

• #79 Codex roots/list protocol compatibility

• #18 dot-directory merge + hooks install migration

Removed -- Dead Code (Stabilization)

• src/store/json-store.ts — JsonBackend class deleted; no runtime or test references. SQLite is sole canonical backend, DegradedBackend is read-only fallback.

• appendArchivedObservations / loadArchivedObservations — dead code removed from persistence-json.ts and persistence.ts re-exports. Archives live in SQLite status='archived' rows.

• 'json' from getBackendName() return type — union narrowed to 'sqlite' | 'degraded' across ObservationStore interface, SqliteBackend, DegradedBackend.

Changed -- Stale Naming Cleanup

• All JsonBackend references in comments and test names replaced with DegradedBackend.

• observations.ts:71 "For JsonBackend" → "For DegradedBackend".

• obs-store.ts:91 "JsonBackend: no-op" → "DegradedBackend: no-op".

• sqlite-store.test.ts:389 test name "falls back to JsonBackend" → "falls back to DegradedBackend", assertion ['sqlite', 'json'] → ['sqlite', 'degraded'].

Fixed -- Dashboard Semantics Closure

• Mode/Port semantics — DashboardState now includes mode ('standalone' | 'control-plane') and port. /api/project returns mode, port, and mcpEndpoint. Dashboard UI shows mode banner with i18n support.

• Team API contract — /api/team in control-plane mode now normalizes SQLite snake_case rows to camelCase, matching the frontend contract. Added listAllAgents, listAllLocks, listAllTasks to TeamStore for global scope.

• Team page crash-proof — loadTeam() safely normalizes missing/null fields (a.id, tk.deps, l.lockedBy). No more Cannot read properties of undefined errors.

• Project switcher sync — Switching project in the dashboard now calls /api/set-current-project to sync backend state.

• i18n coverage — All remaining hardcoded English strings in Team page (time ago, lock TTL, agent time labels, session count) now use t() with Chinese translations.

• CLI help alignment — memorix dashboard and memorix serve-http JSDoc now document mode semantics (Standalone=3210, Control Plane=3211). Startup logs show mode label.

Changed -- Test Suite Stabilization

• E2e demo tests (tests/e2e/) excluded from default vitest run — these test CLI-agent demo artifacts, not Memorix product code. Available via npm run test:e2e.

• Live LLM quality tests excluded from default suite — require MEMORIX_RUN_LIVE_LLM_TESTS=1. Rules-only fallback test preserved in tests/memory/formation-rules-fallback.test.ts. Available via npm run test:llm-live.

• Coordinator merge-conflict test made deterministic — synchronous conflict file writes in spawn() instead of async setTimeout race. Removed { retry: 3 }.

• Default test suite: 156 files, 2064 tests, 0 failed.

Added -- Hooks Test Coverage

• Audit ledger lost/corrupted → re-install recovers audit entry (codex).

• Non-shared-rules agent install/uninstall (claude) verifying config file creation and cleanup.

[1.0.6] - 2026-04-05

Added -- Memory Provenance and Layered Retrieval

• Provenance foundation -- Observations now carry sourceDetail (explicit / hook / git-ingest) and valueCategory (core / contextual / ephemeral). All ten write-path call sites annotated. Backward-compatible: old data without new fields parses cleanly.

• Layered disclosure (L1/L2/L3) -- memorix_session_start now separates routing hints (L1), working context (L2), and deep evidence (L3). Session injection scores observations by source and value category so hook noise stays out of working context.

• Evidence retrieval -- memorix_detail and memorix_timeline now surface provenance cues (source badge, evidence basis) so operators can trace why a memory exists and what supports it.

• Verification-aware evidence -- Detail and timeline outputs distinguish direct, summarized, derived, and repository-backed evidence without requiring a full citation framework.

• Citation-lite -- Evidence-bearing surfaces emit lightweight citation hints ([source: git], [verified: repo-backed]) to support "why surfaced" and "what supports this" queries.

• Retrieval tuning -- Source-aware boost treats git-ingest as first-class git evidence for intent-aware ranking. Lightweight provenance tiebreaking for ambiguous retrieval results. L1 routing surfaces active entities as next-hop search guidance.

• Graph routing hints -- Knowledge graph neighborhood is used for lightweight retrieval enrichment and entity-affinity scoring without a full graph rewrite.

Added -- Task-Line Scoping, Secret Safety, and Attribution Hardening

• Task-line scoping -- Search and session context now bias toward the current entity/task-line/subdomain, reducing cross-workstream contamination within a single project bucket.

• Secret safety -- Store-time detection blocks obvious credentials, passwords, and tokens from entering durable memory. Retrieval-time redaction acts as a second safety net for already-stored sensitive data.

• Project attribution hardening -- Write-path consistency checks reduce wrong-bucket writes. memorix_audit_project scans for misattributed observations and reports them with confidence levels.

• Ambiguous-target attribution fix -- Observations stored during ambiguous project context are now flagged rather than silently written to the wrong bucket.

Added -- Retention, Cleanup, and Operator Remediation

• Retention calibration -- Source-aware retention multipliers (hook 0.5x, git-ingest 1.5x) and value-category multipliers (ephemeral 0.5x, core 2.0x) with a 7-day minimum floor. Immunity refined: only critical importance and core valueCategory grant permanent immunity; high-importance types keep long retention but can now decay.

• Retention explainability -- memorix_retention action="stale" shows a full table with per-observation retention explanation (importance, multipliers, effective days, zone, immunity reason).

• Cleanup remediation loop -- memorix_retention (stale/report), memorix_audit_project, and memorix_resolve now form a coherent operator loop. Each output includes structured Suggested IDs: [...] blocks and explicit next-step guidance. memorix_resolve links back to retention report for closed-loop cleanup.

Added -- OpenCode Plugin Improvements

• post_compact event -- New post_compact hook event type. OpenCode's session.compacted event correctly maps to post_compact (was incorrectly mapped to pre_compact). Plugin event handler triggers runHook side-effect on compaction completion.

• Structured compaction prompt -- OpenCode compaction prompt rewritten as a structured continuation format requesting current task, key decisions, active files, blockers, next steps, active entities, and memorix context. No longer promises automatic memorix_store / memorix_session_start invocation during compaction.

Fixed

• #45 OpenCode compaction -- Compaction prompt no longer makes misleading tool-call promises. session.compacted event now fires a real side-effect via runHook. Normalizer mapping corrected to post_compact.

• #46 Dotenv load order -- loadDotenv() now runs before getEmbeddingProvider() in status, doctor, and TUI entry points, fixing "No API key" errors when .env credentials were present.

• #48 Ingest log dedup -- memorix ingest log now deduplicates by commit hash, matching the behavior of ingest commit and TUI batch ingest. Repeated runs skip already-ingested commits.

Stats

• Tests: 1439 passed | 2 skipped (102 files)

• Phases landed: 11 (provenance -> layered disclosure -> evidence -> verification -> citation-lite -> retrieval tuning -> graph routing -> task-line/secret -> attribution -> retention -> cleanup ergonomics)

[1.0.5] - 2026-03-24

Added

• TUI workbench matured into a product-grade terminal UI - Added an Ink-native /configure flow, interactive sidebar navigation, unified keyboard model, better no-project empty state, compact responsive layouts, and broader TUI interaction coverage.

• Gemini CLI as a first-class integration target - Added a dedicated Gemini CLI target across TUI integrate flows, workspace adapters, rules sync, hook normalization, and MCP config generation.

• Release-blocker regression suite - Added real embedded serve-http route tests for CORS and /api/config, plus cold-start CLI search regression coverage against persisted observations.

• Silent auto-update wiring - Wired the existing updater into real runtime entry points so TUI and HTTP control-plane starts can background-check and silently install newer npm releases without blocking startup.

Changed

• Control plane stability and scope semantics - Hardened HTTP project binding, dashboard API behavior, project-scoped health/search diagnostics, and release-readiness around multi-project sessions.

• Product positioning and integration messaging - Updated README, AI-facing docs, and agent/rules entry docs to foreground Memorix as an open-source cross-agent memory layer compatible with ten major coding IDEs and MCP hosts.

• Search and retrieval transparency - Search mode reporting is now project-scoped end-to-end, including TUI, embedded stats, and MCP search responses.

• Session handoff semantics - memorix_session_start now separates Recent Handoff, Key Project Memories, and Recent Session History so recency-first handoff context is no longer mixed with long-term importance-ranked memories.

Fixed

• Embedded dashboard security and config isolation - Fixed localhost-only CORS behavior for embedded dashboard JSON APIs and closed the /api/config?project=... startup-project YAML leak.

• Cross-project retrieval correctness - Fixed memorix_detail bare numeric IDs to remain project-safe instead of opening observations from another project.

• Concurrent memory write consistency - Fixed topicKey upsert races by rechecking authoritative disk state under the file lock before deciding whether to create or update.

• CLI cold-start search regression - Fixed memorix search so persisted observations are hydrated into the Orama index on a fresh process before searching.

• Embedding provider resilience - Fixed API embedding batch-limit handling with provider-aware chunking, automatic split-and-retry on oversized batches, and retry handling for transient 429/5xx errors.

• OpenCode stale plugin detection - Added generated-version markers and hook-status detection so outdated OpenCode plugin installs are surfaced and can be reinstalled before they corrupt the TUI experience.

• Documentation encoding regressions - Restored clean UTF-8 copy for Chinese README content and agent/rules entry docs so release docs match the current product shape.

Known Limitations

• Gemini CLI / Antigravity shared .gemini/* ecosystem - This follows the official Gemini ecosystem design. Integrations are independent at the target/adapter level, but hook runtime identity can still behave as "last installer wins" because both share the same official hook config surface.

Stats

• Tests: 1099/1101 passing (82 files, 2 skipped)

• Runtime surfaces covered before release: stdio MCP, HTTP control plane, dashboard, TUI workbench, silent auto-update, Gemini CLI integration, git-hook ingest, and cold-start CLI search

[1.0.4] -- 2026-03-17

Added

• Git Memory pipeline -- git commit can now flow directly into Memorix via memorix git-hook, memorix git-hook-uninstall, and memorix ingest commit --auto. Stored observations now carry source and commitHash, and Git memories can be filtered explicitly with source: "git".

• Reasoning Memory tools -- Added memorix_store_reasoning and memorix_search_reasoning so design rationale, alternatives, constraints, and risks can be stored and searched as a first-class memory layer.

• Source-aware retrieval and cross-linking -- Search now boosts Git, reasoning, and problem-solution memories differently based on query intent. Git memories and reasoning memories can cross-reference each other via related commits and shared entities.

• Structured config model -- Added project/user memorix.yml, project/user .env loading, memorix init, and configuration provenance diagnostics in memorix status.

• Dashboard control plane upgrades -- Added Git Memory, Config Provenance, and Identity Health views, plus richer stats and a stabilized graph layout for the HTTP dashboard.

Changed

• Documentation consolidation -- Reworked README, README.zh-CN, setup, architecture, API reference, configuration, Git Memory, and development guides so they match the current product model: local-first platform, memorix.yml + .env, Git Memory, HTTP dashboard, and the four-layer architecture.

• Project detection model -- Project identity now centers on real Git roots, MCP roots support, system-directory fallback handling, and runtime project switching instead of older placeholder-style fallback identities.

• Dashboard usage model -- memorix background start is now the primary long-lived HTTP control-plane entrypoint when you want HTTP transport, collaboration features, and dashboard access in one place. memorix serve-http --port 3211 remains the foreground/debug variant.

Fixed

• Project identity drift -- Fixed Codex/Windsurf startup issues that produced local/System32, IDE-installation-directory identities, or other incorrect local project bindings.

• Worktree-safe Git hooks -- Hook installation, uninstall, auto-install checks, and status reporting now resolve hooks directories correctly for both normal repos and Git worktrees.

• Runtime config correctness -- Fixed project-level memorix.yml not reaching runtime getters, .env values leaking across project switches, and legacy config.json not showing up correctly in provenance diagnostics.

• Git Memory quality -- Added noise filtering, preserved release/version milestone commits, and implemented memorix ingest commit --force as an escape hatch for manual ingestion.

• Cross-project detail retrieval -- Global search results can now be opened reliably with project-aware refs instead of colliding on observation IDs from different projects.

• Skill generation noise -- memorix_skills generate now filters low-signal command-history observations like git, gh, npm, and npx so generated skills stay project-relevant.

• OpenCode static plugin noise -- Merged the first external PR to silence console.log spam in the static OpenCode plugin without reintroducing session lifecycle side effects.

• CI/publish flow -- Restored CI green after type/test regressions and changed npm publishing workflow to manual trigger instead of automatic release publishing.

Stats

• Tests: 879/879 passing across 68 files

• Runtime modes: stdio MCP (memorix serve), HTTP MCP + dashboard (memorix background start by default, or memorix serve-http --port 3211 in the foreground), and standalone dashboard remain supported

[1.0.3] -- 2026-03-14

Added

• Memory Formation Pipeline -- Three-stage pipeline (Extract -> Resolve -> Evaluate) runs in shadow mode on every memorix_store call and hooks trigger. Collects quality metrics without affecting storage decisions.

  • Extract: Automatic fact extraction from narratives, title normalization, entity resolution against Knowledge Graph, observation type verification.

  • Resolve: 4 resolution actions (new/merge/evolve/discard) based on similarity scoring, word overlap, and contradiction detection.

  • Evaluate: Multi-factor knowledge value assessment (type weight, fact density, specificity, causal reasoning, noise detection). Categorizes memories as core/contextual/ephemeral.

• memorix_formation_metrics tool -- New MCP tool to query aggregated Formation Pipeline metrics (value scores, resolution actions, extraction rates, processing times).

• getEntityNames() method on KnowledgeGraphManager for Formation Pipeline entity resolution.

Stats

• Default MCP Tools: 23 (+1: memorix_formation_metrics)

• Tests: 803/803 passing across 60 files (+50 new Formation Pipeline tests)

• Hooks safety: handler.ts +21 lines (shadow call only), zero modification to existing hook logic

[1.0.2] -- 2026-03-14

Fixed

• MCP Server version mismatch -- Server now reports the correct version from package.json (was hardcoded 0.1.0). Injected at build time via tsup define.

• CI Node.js matrix -- Removed Node 18 from CI matrix to match engines: >=20 in package.json.

• Orama reindex idempotency -- reindexObservations() now resets the Orama DB before rebuilding, eliminating "document already exists" errors in multi-session scenarios.

• E2E tests no longer touch real user data -- Mini-skills E2E tests now use a temporary directory with synthetic observations instead of reading/writing ~/.memorix/data/.

[1.0.1] -- 2026-03-14

Fixed

• OpenCode stdout pollution -- Removed all console.log / console.error from the generated OpenCode plugin template. Hooks now run fully silent. (fixes #15)

• OpenCode session_id missing -- normalizeOpenCode() now reads session_id from the payload instead of hardcoding empty string. Plugin template generates and injects a stable session ID per plugin lifetime. (fixes #14)

• Auto-install hooks scope -- Hooks are now only auto-installed for IDEs whose project-level config directory already exists (e.g., .cursor/, .windsurf/), preventing unwanted IDE directories from appearing in projects opened with a different IDE.

Added

• MEMORIX_DATA_DIR environment variable -- Override the default data directory (~/.memorix/data/) by setting MEMORIX_DATA_DIR. Applied consistently across persistence, alias registry, and embedding cache.

[1.0.0] -- 2026-03-09

🎉 First Stable Release

Memorix reaches v1.0.0 -- all major features complete. Future versions will iterate based on AI/agent ecosystem evolution.

Added

• Multi-Agent Team Collaboration -- 4 team tools (team_manage, team_file_lock, team_task, team_message) for cross-IDE agent coordination. File-based persistence via team-state.json. Verified: Windsurf <-> Antigravity bidirectional communication.

• Auto-Cleanup on Startup -- Background retention archiving and intelligent deduplication run automatically in deferredInit. With LLM configured: semantic dedup via any OpenAI-compatible model. Without LLM: Jaccard similarity consolidation. Zero manual maintenance required.

• memorix_transfer tool -- Merged memorix_export + memorix_import into a single tool with action: "export" | "import".

• TEAM.md -- Multi-agent coordination protocol documentation.

Changed

• Tool consolidation: 41 -> 22 default tools (-46%)

  • Team tools: 13 individual -> 4 merged (action parameter pattern)

  • Knowledge Graph tools: 9 -> conditional via ~/.memorix/settings.json ({ "knowledgeGraph": true })

  • Export+Import: 2 -> 1 (memorix_transfer)

• Dashboard Team Panel -- Redesigned with Iconify icons, Material Design 3 style. Agent cards, task lists, message panel, file lock panel.

• README updated for v1.0.0 stable (EN + 中文).

Fixed

• Windows EPERM file lock race condition -- Treat EPERM same as EEXIST in file-lock.ts.

• PowerShell BOM in config.json -- Set-Content -Encoding UTF8 adds BOM in PS 5.x, breaking JSON.parse. Always use Node.js for config file writes.

Production Hardening

• Cross-session shared team state

• Inbox capped at 200 messages with auto-eviction

• Session timeout GC (30min idle -> auto-close)

• Send to inactive agent rejected

• Agent leave releases file locks + clears inbox

• Orphaned task rescue when assignee inactive

• Input validation: agent name max 100, message max 10KB

Stats

• Default MCP Tools: 22 (+9 optional KG)

• Tests: 753/753 passing across 56 files

• IDE Support: 10 agents (Cursor, Windsurf, Claude Code, Codex, Copilot, Kiro, Antigravity, OpenCode, Trae, Gemini CLI)

[0.12.0] -- 2026-03-08

Added

• Intent-Aware Recall -- Search understands query intent ("why X?" prioritizes decisions/trade-offs, "how to X?" prioritizes how-it-works).

• MCP Deadlock Fix -- Resolved stdio transport deadlock under high concurrency.

• Dashboard Dark Theme Fix -- Proper dark mode support across all panels.

• Build Race Condition Fix -- Fixed tsup parallel build race condition.

[0.11.0] -- 2026-03-07

Added

• Mini-Skills (memorix_promote) -- Promote observations to permanent skills that auto-inject at session start. Never decay, project-scoped.

• LLM Quality Engine -- Compact-on-write (duplicate detection at write time), narrative compression (~27% token reduction), search reranking (60% queries improved).

• memorix_deduplicate tool -- LLM-powered semantic deduplication with dry-run support.

• memorix_resolve tool -- Mark completed tasks and fixed bugs as resolved to prevent context pollution.

Fixed

• Retention decay fix -- Reclassified what-changed/discovery to low retention (30d instead of 90d).

Stats

• Tests: 641 -> 674 passing

[0.10.6] -- 2026-03-06

Fixed

• Minor stability improvements.

[0.10.5] -- 2026-03-05

Fixed

• 🔴 Critical: Antigravity MCP connection failure -- CLI banner (starting with 🧠 emoji, UTF-8 F0 9F A7 A0) was written to stdout via console.log in the non-interactive branch. When citty dispatches to serve subcommand, it calls parent run() first, polluting the MCP JSON-RPC stream. Go's encoding/json in Antigravity failed on the first byte 0xF0 with invalid character 'ð'. Fix: console.log -> console.error for all CLI banner output.

• 🔴 Critical: Claude Code Stop hook schema validation failure -- hookSpecificOutput was returned for all hook events, but Claude Code only supports it for PreToolUse, UserPromptSubmit, and PostToolUse. Events like SessionStart, Stop, and PreCompact with hookSpecificOutput triggered JSON validation failed: Invalid input. Fix: only include hookSpecificOutput for the 3 supported event types.

• Claude Code hook_event_name not read -- Handler read payload.hookEventName (camelCase) but Claude Code sends hook_event_name (snake_case), causing hookEventName to always be empty and hookSpecificOutput to be {}.

• Windows hook stdin piping broken -- cmd /c memorix hook wrapper broke stdin piping for hook event JSON. Changed to memorix.cmd hook which directly invokes the CMD shim and properly forwards stdin.

• CLI emoji removed -- All emoji in CLI output replaced with plain text markers ([OK], [FAIL], [WARN], [SKIP], [DRY RUN]) for enterprise-grade compatibility and to prevent future UTF-8 encoding issues.

[0.9.25] -- 2026-02-28

Fixed

• Windsurf "no tools returned" -- Transport-first architecture caused Windsurf to query tools/list before tools were registered. Normal path now registers tools first, then connects transport. Roots path (invalid cwd) still connects first to query listRoots.

• Windsurf rules not activated -- Generated .windsurf/rules/memorix.md lacked YAML frontmatter (trigger: always_on). Windsurf ignored the file without it. Also added alwaysApply: true frontmatter for Cursor .mdc files.

• Windsurf hook post_command content too short -- Normalizer didn't extract commandOutput from Windsurf post_command events, causing content to be <30 chars and filtered out.

• Hook hot-reload broken on Windows -- fs.watch() lost track of observations.json after atomicWriteFile (which uses rename()). Switched to fs.watchFile with 2s polling for reliable cross-platform hot-reload. Hook-written memories are now searchable within ~4 seconds.

[0.9.18] -- 2026-02-26

Fixed

• Self-referential command noise -- Bash commands that inspect memorix's own data (e.g. node -e "...observations.json...", cat ~/.memorix/...) were being stored as observations, creating a feedback loop. Now filtered alongside memorix_internal tools.

[0.9.17] -- 2026-02-26

Fixed

• Session activity noise -- Empty session_end events were unconditionally stored, generating ~8.5% of all observations as useless "Session activity (discovery)" entries. Now requires content ≥ 50 chars, matching the quality-first philosophy of 0.9.16.

[0.9.16] -- 2026-02-26

Architecture

• Classify -> Policy -> Store pipeline -- Replaced the monolithic switch/case handler (527 lines) with a clean declarative pipeline (432 lines). Inspired by claude-mem's store-first philosophy and mcp-memory-service's configurable scoring.

• Tool Taxonomy -- classifyTool() categorizes tools into file_modify | file_read | command | search | memorix_internal | unknown. Each category has a declarative StoragePolicy (store mode, minLength, defaultType).

• Pattern detection = classification only -- Pattern detection now only determines observation type (decision, error, etc.), not whether to store. Storage decisions are made by policy.

• Unified TYPE_EMOJI -- Single exported constant, eliminating 3 duplicated copies across handler and session_start.