mempal
Memory framework
First committed by ZhangHanDong
Published in ZhangHanDong/mempal. Publisher page
284 GitHub stars
What it is
AI Agent Memory
AutoGineer overview, from the brain’s own files
Mempal is a project memory system designed for coding agents, implemented as a single binary. It enables agents to store, search, and retrieve decision context across sessions. The system utilizes a hybrid search approach combining BM25 keyword matching and vector semantic search, alongside a knowledge graph for tracking entity relationships. It includes a self-describing protocol that allows agents to understand how to interact with the memory system without manual system prompt configuration. Mempal supports cross-project tunnels, multilingual embedding, and collaborative agent workflows through inbox-based handoffs.
How it works
Memory model: Hybrid
AutoGineer overview, from the brain’s own files
Mempal stores data in a SQLite database located at ~/.mempal/palace.db. It ingests project information by chunking and embedding content, which is then searchable via CLI or MCP tools. The system uses a memory protocol to guide agent behavior, such as verifying facts before ingestion and citing sources. It supports collaborative agent interactions by allowing agents to push messages to a per-project inbox, which are then processed via hooks during user prompt submission. The knowledge graph tracks triples with temporal validity, and the system provides tools for fact-checking and context assembly.
| Folder | Role | Files |
|---|---|---|
| .claude | agent-config | 2 |
| .github | agent-config | 3 |
| docs | docs | 129 |
| hooks | hooks | 2 |
| src | source-code | 30 |
| tests | tests | 31 |
| benchmarks | folder | 1 |
| books | folder | 41 |
| crates | folder | 71 |
| specs | folder | 131 |
The agent writes
- record_checked evidence: AGENTS.md
Requirements
Runtimes: Rust
Databases: SQLite
Works with: Claude Code, OpenAI Codex, GitHub Copilot, Cursor, Kiro, OpenCode, Amp, Cline, JetBrains Junie, Devin (ex-Windsurf)
Set it up
This brain runs code automatically: 2 hooks (claude-code). Review .claude/settings.json before opening the folder in Claude Code. No step below turns a hook on for you.
.claude/settings.json:if echo "$TOOL_INPUT" | grep -q 'git commit'; then echo 'MEMPAL REMINDER: You just committed code. Call mempal_ingest to save the decision context (what was built, why, what was considered). Rule 4: SAVE AFTER DECISIONS.'; fi.claude/settings.json:bash .claude/hooks/user-prompt-submit.sh
- Not scanned by AutoGineerThe author’s MCP configuration (.mcp.json) also starts it in a way AutoGineer does not offer. Read that file: AutoGineer did not scan that path; what it runs is listed under “This brain runs”. https://github.com/ZhangHanDong/mempal/blob/9ed705d78321daf5878e9579777b1a471e97c6f1/.mcp.json
- Not scanned by AutoGineerThe author’s README also describes installing it with cargo. Read its install section: AutoGineer did not scan that path. https://github.com/ZhangHanDong/mempal/blob/9ed705d78321daf5878e9579777b1a471e97c6f1/README.md#quick-start
First committed by ZhangHanDong. Published in ZhangHanDong/mempal. Licence: MIT License.
Usage
AutoGineer overview, from the brain’s own files
- Initialize project memory for a directory
- Ingest project files into a specific wing
- Perform hybrid search for past decisions
- Fact-check assertions against the knowledge graph
- Collaborate between agents using inbox pushes
- Manage knowledge lifecycle and anchor publication
- “mempal search "auth decision clerk"”
- Returns past decisions related to authentication and Clerk, including source citations.
- “mempal fact-check path/to/file”
- Identifies potential contradictions like similar name conflicts or stale facts against the knowledge graph.
Documentation by ZhangHanDong
mempal
Image: Crates.io Image: Downloads Image: docs.rs Image: CI Image: Release Image: License: MIT
Project memory for coding agents. Single binary, cargo install mempal, find past decisions with citations in seconds.
Latest release: v0.9.1 (2026-08-19).
What It Does
Agent writes code → commits → mempal saves the decision context
Next session (any agent) → mempal search → finds the decision with source citationHybrid search: BM25 keyword matching + vector semantic search, merged via Reciprocal Rank Fusion
Knowledge graph: subject-predicate-object triples with temporal validity (valid_from/valid_to)
Cross-project tunnels: automatic discovery when the same room appears in multiple wings
Self-describing protocol: MEMORY_PROTOCOL embedded in MCP ServerInfo teaches any agent how to use mempal — no system prompt configuration required
Multilingual: model2vec-rs (BGE-M3 distilled) as default embedder, zero native dependencies
Single file: everything lives in
~/.mempal/palace.db(SQLite + sqlite-vec)
Quick Start
cargo install mempal --version 0.9.1 --locked
mempal init ~/code/myapp
mempal ingest ~/code/myapp --wing myapp
mempal search "auth decision clerk"
mempal wake-upWith REST support:
cargo install mempal --version 0.9.1 --locked --features restFrom a source checkout:
cargo install --path . --lockedConfiguration
Config at ~/.mempal/config.toml (optional, defaults work without it):
db_path = "~/.mempal/palace.db"
[embed]
backend = "model2vec" # default, zero native deps
# model = "minishlab/potion-multilingual-128M" # default multilingual model (1024d)Other backends:
# Local ONNX (requires --features onnx)
[embed]
backend = "onnx"
# External API
[embed]
backend = "api"
api_endpoint = "http://localhost:11434/api/embeddings"
api_model = "nomic-embed-text"
dimensions = 768 # optional; defaults to 384 when omittedCommands
| Command | Purpose |
|---|---|
mempal init <DIR> [--dry-run] | Infer wing/rooms from project tree |
mempal ingest <DIR> --wing <W> [--dry-run] | Chunk, embed, and store |
mempal search <QUERY> [--wing W] [--room R] [--json] | Hybrid search (BM25 + vector + RRF) |
mempal wake-up [--format aaak] | Context refresh, sorted by importance |
mempal compress <TEXT> | AAAK format output |
mempal delete <DRAWER_ID> | Soft-delete a drawer |
mempal purge [--before TIMESTAMP] | Permanently remove soft-deleted drawers |
mempal kg add <S> <P> <O> | Add a knowledge graph triple |
mempal kg query [--subject S] [--predicate P] | Query triples |
mempal kg timeline <ENTITY> | Chronological view of an entity |
mempal kg stats | Knowledge graph statistics |
mempal tunnels | Cross-wing room links |
mempal taxonomy list / edit | Manage routing keywords |
mempal reindex | Re-embed all drawers after model change |
mempal status | DB stats, schema version, scopes |
mempal serve [--mcp] | MCP server (+ REST with feature) |
mempal cowork-install-hooks [--global-codex] | Install UserPromptSubmit hooks for Claude Code (+ optional Codex merge) |
mempal cowork-drain --target <claude|codex> | Drain inbox messages (for hook use; exits 0 on any failure) |
mempal cowork-status --cwd <PATH> | Read-only view of both inboxes at <PATH> |
mempal fact-check [PATH|-] [--wing W] [--room R] [--now <UNIX_SECS>] | Offline contradiction check against KG triples + known entities |
mempal bench longmemeval <FILE> | LongMemEval retrieval benchmark |
Project ignore rules
mempal init <DIR> and directory-mode mempal ingest <DIR> use the target
project's ignore rules by default. They respect .gitignore, .git/info/exclude,
global Git excludes, and a root-level .mempalignore. Built artifacts that should
never become project memory can be listed in .mempalignore without changing the
project's Git behavior.
Use --ignore-file <PATH> to add one or more extra ignore files for a run. Use
--no-gitignore or --no-mempalignore to disable either default source. Explicit
file ingest remains explicit: mempal ingest path/to/file --wing <W> still ingests
that file even if the project ignore rules would skip it during directory traversal.
MCP Server (10 tools)
mempal serve --mcp exposes these tools via Model Context Protocol:
| Tool | Purpose |
|---|---|
mempal_status | State + protocol + AAAK spec (teaches agent on first call) |
mempal_search | Hybrid search with tunnel hints, citations, and AAAK-derived structured signals |
mempal_ingest | Store memories with optional importance (0-5) and dry_run; reports lock_wait_ms when concurrent ingest was observed |
mempal_delete | Soft-delete with audit trail |
mempal_taxonomy | List or edit routing keywords |
mempal_kg | Knowledge graph: add/query/invalidate/timeline/stats |
mempal_tunnels | Cross-wing room discovery |
mempal_peek_partner | Read partner agent's live session (Claude ↔ Codex), pure read, never writes |
mempal_session_peek | Read an explicit Claude/Codex local session by concrete tool + cwd, including same-tool cross-project reads |
mempal_cowork_push | Send a short handoff message to partner agent's inbox (at-next-submit delivery) |
mempal_fact_check | Offline contradiction detection vs KG triples + known entities (similar-name, relation mismatch, stale facts) |
The server embeds MEMORY_PROTOCOL (11 behavioral rules) in the MCP initialize.instructions field. Any MCP client learns the workflow automatically.
Memory Protocol
mempal teaches agents these rules through self-description:
FIRST-TIME SETUP — call
mempal_statusto discover wings before filteringWAKE UP — different clients have different pre-load mechanisms
VERIFY BEFORE ASSERTING — search before stating project facts
QUERY WHEN UNCERTAIN — search on "why did we...", "last time we..." 3a. TRANSLATE TO ENGLISH — translate non-English queries before searching
SAVE AFTER DECISIONS — persist rationale, not just outcomes
CITE EVERYTHING — reference drawer_id and source_file 5a. KEEP A DIARY — record behavioral observations in wing="agent-diary"
PARTNER AWARENESS — use
mempal_peek_partnerfor live partner-agent session,mempal_session_peekfor explicit same-tool/cross-project local sessions, and mempal search only for crystallized drawersDECISION CAPTURE —
mempal_ingestis for firm decisions only; include partner input when peek informed the callCOWORK PUSH — use
mempal_cowork_pushas the SEND primitive in the SEND/READ/PERSIST triad; at-next-submit delivery, not real-timeVERIFY BEFORE INGEST — call
mempal_fact_checkbefore persisting a decision that asserts entity relationships; it catches similar-name typos, relation mismatches against the KG, and stale facts with expiredvalid_to
Agent Cowork (P6 peek + P8 push)
Two coding agents (Claude Code and Codex) can collaborate on the same repo through a per-project inbox + hook-driven injection channel, on top of mempal_peek_partner (read live partner session), mempal_session_peek (read an explicit local session by concrete tool + cwd), and mempal_cowork_push (send ephemeral handoff).
Install hooks once per repo (run at the repo root):
mempal cowork-install-hooks --global-codexThis writes:
.claude/hooks/user-prompt-submit.sh+ merges a registration entry into.claude/settings.jsonso Claude Code fires the hook on every user prompt.~/.codex/hooks.jsonUserPromptSubmit entry so Codex fires the same drain on every user prompt.
The --global-codex part is optional. The re-run is idempotent and self-heals stale/wrong drain entries — re-installing after a mempal upgrade is always safe.
Delivery is at-next-UserPromptSubmit, not real-time: a push from Claude to Codex becomes visible only when the Codex user submits their next prompt, at which point the hook drains the inbox and prepends the message as additionalContext on that turn.
Check inbox state at any time without draining:
mempal cowork-status --cwd "$PWD"Known limitations
Codex feature flag dependency: Codex's hooks runtime is gated behind the
codex_hooksfeature flag (currently "under development" in shippedcodex-cli). If the flag is off, Codex silently ignores~/.codex/hooks.json.install-hooksdetects this and prints a warning with the activation command:codex features enable codex_hooks.Two Claude-side artifacts: Claude Code does not auto-discover hook scripts by filename. Both
.claude/hooks/user-prompt-submit.shand the matching entry in.claude/settings.jsonare required.install-hookswrites both; do not remove either by hand.TUI restart needed after config changes on the Codex side: Codex reads
config.toml+hooks.jsonat process startup only. After enabling the feature flag or runninginstall-hooks, fully quit and relaunch the Codex TUI before expecting hooks to fire.MCP server re-spawn: Claude Code spawns the mempal MCP server at client startup. After upgrading the mempal binary (
cargo install ...), restart Claude Code so the MCP server respawns and exposes newly added tools likemempal_cowork_pushormempal_fact_check.Bidirectional scope:
mempal_cowork_pushcurrently requires an MCP client identifying itself asclaude-codeorcodex(or their aliases). Generic MCP clients cannot push because caller identity is required to fill thefromfield and enforce self-push rejection. This is by design for the Claude ↔ Codex pair.
Concurrent Ingest Safety (P9-B)
Two agents writing to the same source simultaneously used to be a TOCTOU race: both would pass the dedup check, both would insert, producing duplicate drawers or mismatched vectors. Since 0.4.0, mempal_ingest and ingest_file_with_options acquire a per-source advisory lock before entering the dedup + insert critical section.
Lock files live at
~/.mempal/locks/<16-hex>.lock, created lazily, released on guard drop.5 s timeout, 50 ms retry + jitter;
LockError::Timeoutsurfaces as aningesterror.Every non-dry-run response carries
lock_wait_ms: Option<u64>so agents can detect contention.Dry-run does not acquire the lock (no writes, no race).
Unix only in 0.4.0. On Windows the lock path is a no-op fallback;
LockFileExsupport is tracked for a follow-up.
Offline Fact Checking (P9-A)
mempal_fact_check — and its CLI counterpart mempal fact-check — compare a text blob against the existing KG triples + the entity registry derived from recent drawers. It flags three issue classes, all deterministic and zero-LLM:
| Issue | Trigger |
|---|---|
SimilarNameConflict | Text mentions a name within Levenshtein distance ≤ 2 of a known entity, and the names are not equal. |
RelationContradiction | Text asserts a predicate (e.g. brother_of) that's in the incompatibility dictionary against an existing KG triple with the same (subject, object) endpoints. |
StaleFact | Text asserts a triple whose KG row has valid_to < now (Unix seconds). |
Extracted triples today cover three narrow patterns: "X is Y's ROLE", "X works at / for Y", and "X is [the|a|an] ROLE of Y". Unknown sentence shapes are silently ignored, so the tool errs toward under-reporting rather than false positives.
Protocol Rule 11 guides agents to run this before ingesting a decision that asserts entity relationships. See specs/p9-fact-checker.spec.md for the full contract.
Search Architecture
query → BM25 (FTS5) → ranked by keyword match
→ Vector (sqlite-vec) → ranked by semantic similarity
→ RRF Fusion (k=60) → merged ranking
→ Wing/Room filter → scoped results
→ Tunnel hints → cross-project referencesKnowledge Graph
mempal kg add "Kai" "recommends" "Clerk"
mempal kg add "Clerk" "replaced" "Auth0" --source-drawer drawer_xxx
mempal kg timeline "Kai"
mempal kg statsTriples support temporal validity — relationships can be invalidated when they expire.
Agent Diary
Cross-session behavioral learning — agents record observations, lessons, and patterns:
# Search diary entries
mempal search "lesson" --wing agent-diary
mempal search "pattern" --wing agent-diary --room claudeDiary entries use the existing mempal_ingest tool with wing="agent-diary" and room=agent-name. MEMORY_PROTOCOL Rule 5a teaches agents to write diary entries. Integrates with Claude Code's auto-dream for automatic memory consolidation.
Ingest Formats (5)
| Format | Auto-detected by |
|---|---|
| Claude Code JSONL | type + message fields |
| ChatGPT JSON | Array or mapping tree |
| Codex CLI JSONL | session_meta + event_msg entries |
| Slack DM JSON | type: "message" + user + text |
| Plain text | Fallback |
AAAK Compression
Output-only format readable by any LLM without decoding:
mempal compress "Kai recommended Clerk over Auth0 based on pricing and DX"
# V1|manual|compress|1744156800|cli
# 0:KAI+CLK+AUT|kai_clerk_auth0|"Kai recommended Clerk over Auth0..."|★★★★|determ|DECISIONChinese text uses jieba-rs POS tagging for proper word segmentation.
Architecture
| Crate | Responsibility |
|---|---|
mempal | Installable CLI package, REST entrypoint, and legacy facade paths |
mempal-embed | Public Embedder / EmbedderFactory traits, API embedder, model2vec and ONNX implementations |
mempal-search-core | Public FTS5 query escaping and RRF rank-fusion primitives |
mempal-agent-memory | Public agent-memory domain types and anchor helpers |
mempal-mcp-protocol | Public self-describing MEMORY_PROTOCOL text |
mempal-store-sqlite | Public SQLite storage layer: schema, migrations, Database, FTS5, sqlite-vec, KG, tunnels, cards, adoption tables |
mempal-runtime | Public runtime workflows: ingest, search orchestration, context, brief, knowledge, factcheck, projects, cowork, doctor, AAAK |
mempal-mcp-server | Public MCP server adapter and tool wiring for the runtime |
Key design choices:
model2vec-rs default embedder — zero native deps, multilingual (BGE-M3 distilled)
ort (ONNX) available behind
onnxfeature flag for max qualityFTS5 for BM25 keyword search — synced via SQLite triggers
Soft-delete with audit trail —
mempal delete+mempal purgeImportance ranking — drawers have 0-5 importance, wake-up sorts by importance
Semantic dedup — ingest warns (doesn't block) when similar content exists
Public workspace crates — reusable libraries are maintained in this repo and published as versioned crates while
cargo install mempalremains the CLI path. P114 keeps the split coarse: storage, runtime, and MCP server are separate crates, but feature families such as knowledge, cowork, and context remain inmempal-runtime.
Development
cargo test --workspace
cargo test --workspace --all-features
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo fmt --all --checkAfter changing the embedding model or dimensions, re-embed existing drawers:
mempal reindexDocs
Usage guide:
docs/usage.mdAAAK dialect:
docs/aaak-dialect.mdSpecs (internal agent-spec contracts, on GitHub): https://github.com/ZhangHanDong/mempal/tree/main/specs
Plans (internal implementation plans, on GitHub): https://github.com/ZhangHanDong/mempal/tree/main/docs/plans
Benchmark:
benchmarks/longmemeval_s_summary.md— includes the older 384d baseline and the newer model2vec 256d run
Book: MemPalace — Reforging Memory in Rust
mempal 的设计分析和完整技术叙事,收录在《MemPalace: AI 记忆的第一性原理》Part 10(第 26-30 章):
| 章节 | 内容 |
|---|---|
| 第 26 章 | 为什么用 Rust 重铸 — 触发点、重写判断、语言选择 |
| 第 27 章 | 保留了什么、改变了什么 — 5 维度对比 + 架构图 |
| 第 28 章 | 自描述协议 — MEMORY_PROTOCOL、7 条规则、agent 生命周期 |
| 第 29 章 | 多 Agent 协作 — Claude↔Codex 接力、反模式发现、agent 日记 |
| 第 30 章 | 诚实的差距 — benchmark 数据、6 个 gap |
docs/MIND-MODEL-IMPLEMENTATION-ARTICLE.zh-CN.md
从记忆工具到可治理的心智模型
本文解释 mempal 心智模型从最初的“道 / 术 / 器”讨论,到 P54-P59 Phase 3 runtime baseline 已实现部分之间的设计决策、实现边界和后续方向。
1. 问题:记忆不只是召回
多数 agent memory 系统一开始都会采用一个简单前提:agent 见过某段信息, 就把它存起来;以后需要时再检索出来,塞回上下文。
这个前提有用,但对一个长期运行的 coding agent 来说远远不够。
coding agent 需要的不只是旧文本。它还需要知道:哪些旧文本是证据,哪些 是稳定结论,哪些是工作流,哪些是工具命令,哪些已经被反例削弱或推翻。 如果所有记忆都被当成同一种对象保存,系统迟早会失去“观察”和“规律”之间 的边界。
mempal 的心智模型工作就是从这个失败模式出发的。
核心论点是:
记忆应该是一个可治理的知识演化层。先记录原始证据,再蒸馏候选知识, 最后只有通过显式 gate 的知识,才可以成为可信的 runtime 指导。
所以这套系统不是“RAG 加 skills”。RAG 负责检索旧材料,skills 负责编码 过程,而心智模型要治理的是:证据如何变成知识,知识如何影响 skill 和工具 选择,runtime 行为又如何反过来推动系统继续演化。
当前实现已经达到一个具体 baseline:
Stage 1 typed drawers 已经区分 evidence 和 knowledge。
runtime context 已经按
dao_tian -> dao_ren -> shu -> qi组装。knowledge lifecycle 已经支持 distill、gate、promote、demote 和向外发布 anchor。
Phase 2 knowledge cards 已经把“被治理的 belief”和 raw evidence drawer 分离。
card retrieval 和 card-aware context 已经存在,但仍然是 opt-in。
Phase 3 runtime adoption evidence 已经通过 schema v9 和 CLI gates 落地。
因此,下一阶段不应该再回头补 baseline,而应该验证这个 baseline 是否真的 改善 agent 的实际行为。
2. 道、术、器:知识轴,而不是存储目录
最初的概念拆分是 dao / shu / qi。
在 mempal 里,它们的含义是:
dao:治理原则和高层知识。shu:可复用的方法、工作流和过程性知识。qi:具体工具、接口、命令和工具相关用法。
dao 又进一步拆成两层:
dao_tian:跨领域高层原则,数量最少,稳定性最高。dao_ren:领域规律,仍然稳定,但有明确 field 边界。
这个拆分重要,是因为 coding agent 对不同知识的使用方式不同。
dao_tian 应该塑造判断。例如“证据先于断言”或者“只有 promotion 没有
demotion 会造成知识污染”。这类知识不应该多,也不应该频繁变化。
dao_ren 应该塑造领域推理。对一个 Rust 项目来说,它可能是关于 schema
migration、citation 语义、agent runtime 行为边界的稳定规则。
shu 应该塑造行动。它可以影响 workflow 和 skill 选择。例如,在实现一个
功能前,先查看现有 spec,写最小 contract,再用确定性工具验证。
qi 应该绑定执行。它说明具体有哪些 CLI、哪些 MCP tool、哪些 feature flag,
以及某个工具接口具体怎么用。
关键设计决策是:dao / shu / qi 不是 project memory、agent memory 或
skill memory 的替代品。它们是不同坐标轴。
完整坐标应该包括:
domain:这条记忆服务于谁,例如 project、agent、skill、global。tier:它是哪类知识,例如dao_ren、shu、qi。field:它属于哪个主题领域。provenance:它来自哪里,例如 runtime、research、human。anchor:它属于哪个持久化范围,例如 global、repo、worktree。
这个正交关系可以避免几类常见错误:
项目局部 workaround 不应该变成普遍规律。
临时工具行为不应该变成稳定领域原则。
workflow tip 不应该被误认为治理原则。
research summary 不应该直接创建
dao。
因此,设计上把 dao 放在 memory layer。外部工具可以产生 evidence,但
promotion 的所有权属于 memory。
3. Evidence 和 Knowledge 必须是不同对象
当前实现最核心的拆分,是 evidence memory 和 knowledge memory。
Evidence memory 保存“看到了什么”:
原始 research 输出
runtime 观察
人类显式 teaching
具体失败
矛盾
反例
Evidence 可以不一致。这不是 bug,而是预期行为。真实世界本来就会给出冲突 观察,agent 应该能先记录这些观察,而不是急着把它们压平成一个结论。
Knowledge memory 保存“系统当前相信什么”:
工具使用说明
workflow 方法
领域模式
治理原则
Knowledge 的数量应该更少,必须可审计,并且有状态。它可以是 candidate、 promoted、canonical、demoted 或 retired。
这个拆分回答了一个基础问题:什么才算真正学习?
学习不是“存了更多文本”。学习发生在观察持续积累、模式被蒸馏、有用候选被 提升、错误或过时 belief 可以被降级的时候。
因此 lifecycle 被拆成四个基本动作:
record:存原始证据。distill:从证据创建候选知识。promote:让知识进入可信 runtime 使用。demote:当知识被反例削弱或过时时,降低或退休它。
这是刻意不对称的设计。Evidence 应该增长快,law 应该增长慢。
项目已经用两种形态实现了这个拆分:
Stage 1 用 typed drawers 作为 bootstrap 模型。
Phase 2 引入专门的 knowledge card tables。
这个演进顺序很重要。它让 mempal 不需要一开始就重写存储系统,而是先用低成 本方式验证模型;等治理形状稳定之后,再抽取出更干净的 knowledge object。
4. Stage 1:Typed Drawers 作为 Bootstrap 架构
在心智模型工作之前,drawer 基本上就是 durable memory unit。Stage 1 没有 替换 drawer 系统,而是让 drawer 具备类型语义。
核心区分变成:
evidence drawer
knowledge drawer
两者共享基础 drawer 字段,但元数据和 runtime 语义不同。
evidence drawer 记录 source-backed material。它包含
memory_kind=evidence、domain、field、provenance、anchor_kind、
anchor_id 等字段。它不携带 tier、status、statement 或 lifecycle refs。
knowledge drawer 记录被蒸馏出的 belief。它包含 memory_kind=knowledge、
statement、tier、status、role-separated evidence refs、scope
constraints、trigger hints 和 anchor metadata。
statement 字段很关键。它是短的 wake-up proposition。较长的 content
用于解释 rationale、examples 和 boundaries。runtime context 应该能先唤醒
一个短结论,而不是每次都把长篇解释塞进上下文。
Stage 1 还引入了 role-separated refs:
supporting_refsverification_refscounterexample_refsteaching_refs
这不只是元数据更整洁。它让 promotion gate 能理解一个 claim 为什么可信。 supporting ref 和 counterexample ref 不能被混在同一个匿名 evidence list 里。
随后实现了最小 lifecycle surface:
mempal knowledge distillmempal knowledge gatemempal knowledge promotemempal knowledge demotemempal knowledge publish-anchor
runtime agent 需要直接访问的部分,也补了 MCP surface:
mempal_knowledge_distillmempal_knowledge_gatemempal_knowledge_promotemempal_knowledge_demotemempal_knowledge_publish_anchor
关键约束是:这些操作不允许调用者直接声明 confidence。Promotion 依赖确定性
policy。例如 Stage 1 policy 要求证据数量、verification refs,并且对
dao_tian 要求人类 reviewer。
这就是系统避免危险捷径的方式:
“agent 发现了一个有趣模式,所以系统学会了一条 law。”
这条捷径被明确禁止。
5. Anchors:Worktree、Repo、Global
心智模型还必须解决项目身份问题。
最早容易想到的是用 wing 表示项目身份。但这不对,因为 wing 是语义分区。
它回答“这条 memory 属于哪个语义区域”,而不是“它来自哪个 checkout”。
最终实现的 anchor model 区分:
globalrepoworktree
这个区分对 coding 工作很必要。
只有 repo anchor 太宽。不同 branch experiment 会互相污染,一个 worktree 中的 失败实验可能影响另一个 worktree 的推理。
只有 worktree anchor 又太窄。稳定项目知识会在不同 checkout 之间碎片化,每个 新 worktree 都会过于空白。
因此设计采用双层 scope:
本地观察先写到
worktree被验证的项目知识再向外发布到
repo只有跨项目原则才发布到
global
这也产生了两种不同的上升运动:
tier promotion,例如
qi -> shu或shu -> dao_renanchor publication,例如
worktree -> repo -> global
两者不是同一个操作。一个 workflow 可能在当前 worktree 内可信,但还不应该在
整个 repo 中共享。一个 repo-level pattern 可能对 mempal 稳定,但还没资格成为
global dao_tian。
runtime context assembly 的 anchor 优先级是:
当前 worktree
当前 repo
global
这样 agent 会先拿到 branch-local context,再拿到稳定项目 memory,最后拿到 通用 law。tier order 仍然在 anchor 逻辑内部生效。
6. Runtime Context:Wake-Up 不是 Context Assembly
实现里一个重要边界,是 wake-up 和 mind-model context 的分离。
Wake-up 仍然是 L0/L1 memory refresh surface。它可以把旧材料重新带回注意力, 但它不负责组装 typed operating guidance。
typed operating guidance 属于:
mempal contextmempal_context
这些 surface 按以下顺序组装 context:
dao_tiandao_renshuqioptional evidence
这个顺序是刻意设计的。
dao_tian 校准世界观。dao_ren 校准当前领域。shu 影响 workflow 和 skill
selection。qi 绑定具体工具。Evidence 用于 grounding、异常处理和证明。
实现上还让 dao_tian 默认保持稀疏。mempal context 和 mempal_context
默认最多注入一条 dao_tian,除非调用者显式提高 budget。这避免了实践执行被
抽象原则淹没。
runtime skill guidance 也被设计成非权威层。trigger_hints 可以 bias workflow
和 tool choice,但不能执行 skill,也不能覆盖 system、user、repo 或 client-native
skill rules。
这个边界非常关键。如果 memory hints 可以直接触发工具或覆盖显式指令,系统就会 变成一个不受治理的 policy engine。当前设计让 memory 只提供指导,agent 仍然 服从正常的 instruction hierarchy。
7. Phase 2:Knowledge Cards 作为被治理的 Beliefs
Typed drawers 证明了模型,但它仍然是 bootstrap 架构。Drawer 适合保存 raw evidence,但不是 governed belief 的最终形态。
Phase 2 在同一个 SQLite palace.db 中引入了专门的 knowledge card storage。
核心表是:
knowledge_cardsknowledge_evidence_linksknowledge_events
这是一个明确的存储决策。Cards 默认不应该放到外部数据库。mempal 的产品不变量 是 local、single-binary、single-file memory palace。把 cards 放在同一个 SQLite 数据库里,可以保持 lifecycle update 的事务性,也避免在模型还没证明需要独立 扩展之前引入额外运维复杂度。
分层关系很清楚:
drawers 仍然是 raw evidence 和 citation root
cards 保存 distilled beliefs
evidence links 把 cards 连接到 source-backed drawers
events 记录 belief 如何演化
一张 card 包含 statement、content、tier、status、domain、field、anchor metadata、trigger hints 和 timestamps。
一条 evidence link 用 role 把 card 连接到 evidence drawer:
supporting
verification
counterexample
teaching
一条 event 记录 lifecycle history:
created
promoted
demoted
retired
linked
unlinked
updated
published_anchor
这让 Phase 2 比 Stage 1 更显式。Drawer 继续保存 evidence,card 保存 belief, event stream 解释这个 belief 为什么发生变化。
已实现的 surface 包括:
mempal knowledge-card createmempal knowledge-card getmempal knowledge-card listmempal knowledge-card linkmempal knowledge-card eventmempal knowledge-card eventsmempal knowledge-card backfill-planmempal knowledge-card backfill-applymempal knowledge-card gatemempal knowledge-card promotemempal knowledge-card demotemempal knowledge-card retrieve
MCP tool mempal_knowledge_cards 也暴露了 read、lifecycle 和 retrieve actions。
关键边界是:Phase 2 cards 已经是 governed objects,但还不是默认 runtime
sources。它们可以被 retrieve,也可以通过 --include-cards 或 include_cards=true
注入 context。但普通 mempal_search 仍然 drawer-based,默认 context 仍然
drawer-only。
这不是实现缺口,而是 trust boundary。
默认 runtime context 是高信任路径。要从 drawer-backed knowledge 切换到 card-backed guidance,必须有 runtime evidence 证明这种切换能改善行为,并且不 破坏 citations、lifecycle semantics 或 context budget。
8. Research 和 Evaluator:证据生产者,不是权威
设计也明确了 external research 和 evaluator 的角色。
research-rs 是外部工具。它是 qi,不是 dao。
它的职责是收集和组织外部材料:
raw sources
wiki pages
schema
index
logs
lint reports
它的输出可以成为 evidence。它可以产生 structured summaries、candidate
insights 和 contradiction signals。但它不能直接创建 dao_tian、promoted
knowledge 或 canonical knowledge。
这个边界已经进入实现 policy:
research output 进入 evidence
candidate knowledge 必须从 evidence refs distill
promotion 仍由 lifecycle gates 控制
contradictions 变成 counterexamples 或 demotion evidence
Evaluator 也有类似边界。
Evaluator 可以建议 promotion 或 demotion。它可以提出 supporting refs、 verification refs、counterexample refs 和 risk notes。它可以解释为什么某个 knowledge item 看起来 ready 或 unsafe。
但 evaluator 不能直接 mutate lifecycle state。
Evaluator output 是 advisory-only。它不能满足 human reviewer requirement,
不能绕过 deterministic gates,也不能 canonicalize dao_tian。
这就是 assistance 和 authority 的区别。系统可以使用 evaluator advice,但不让 evaluator 成为 lifecycle actor。
9. Phase 3 到目前为止:先收集 Runtime Evidence,再增加 Runtime Power
P51 之后,baseline 已经关闭。P52 定义 Phase 3 是新阶段,而不是继续补 P12-P50 没做完的部分。P53 审计 candidate tracks 后确认:没有任何候选方向应该 直接进入默认行为变更。
被推荐的第一个 Phase 3 track 是 runtime adoption evidence。
这个推荐已经通过 P54-P59 实现。
P54 增加 schema v9 表 runtime_adoption_events。这张表记录具体 runtime
signals:
signal 属于哪个 track
发生了哪种 signal
涉及哪个 feature
optional query、context hash、card id、evaluator id、research report id
note、metadata、timestamp
支持的 tracks 包括 runtime adoption、card context、card embedding、evaluator 和 research adapter。支持的 signals 包括 used、accepted、rejected、miss、 rollback、contradiction、neutral。
P55 通过 CLI 暴露这张表:
mempal phase3 adoption recordmempal phase3 adoption listmempal phase3 adoption stats
P56 增加 card context default 的 read-only gate。当前 policy 仍然是 opt-in。
这个 gate 只报告是否有足够 accepted card_context evidence,以及是否没有
rollback signals。
P57 增加 card embedding 的 read-only gate。它明确不增加 card vector schema。 Card embeddings 需要重复、可测量的 miss evidence,证明 linked-evidence retrieval 因为只能匹配 evidence wording 而漏掉了应该通过 card statement 找到 的 active card。
P58 增加 evaluator API 的 read-only gate。它保持 advisory-only 边界。 Evaluator signals 不会 mutate lifecycle state,也不会绕过 gates。
P59 增加 mempal phase3 research-validate-plan。它验证 external research
report contract,包括 report_id、title、sources、findings 和 optional
candidate_insights。它只做 validate 和 plan,不 ingest promoted 或 canonical
knowledge。
模式是一致的:
在增加 runtime power 之前,先收集 runtime evidence。
这就是 Phase 3 的目标。
10. 当前到底已经构建了什么
现在可以把已实现系统概括成一个 governed cognition stack。
存储与检索:
SQLite
palace.db是唯一 local storage root。Drawers 仍然是 raw memory 和 evidence citation roots。
BM25 加 vector retrieval 仍然是基础 search path。
AAAK 仍然是 output formatter,不是 storage dependency。
Stage 1 mind model:
evidence 和 knowledge drawers 已经 typed
statement用于 knowledge wake-updao_tian、dao_ren、shu、qi有明确 status rulesanchors 区分 worktree、repo、global scope
lifecycle operations 是显式且 evidence-backed 的
Runtime context:
mempal context和mempal_context组装 typed guidancecontext order 是
dao_tian -> dao_ren -> shu -> qievidence 是 opt-in
card context 是 opt-in
memory hints bias skill/tool choice,但不执行
Phase 2 cards:
knowledge_cards保存 governed beliefsknowledge_evidence_links保留 role-separated citationsknowledge_events提供 append-only lifecycle auditcards 可以从 Stage 1 knowledge backfill
cards 可以 gate、promote、demote、retrieve
默认 search 不返回 cards
Policy boundaries:
research output 不能直接创建
daoevaluator output 是 advisory-only
dao_tiancanonicalization 需要 human reviewcard embeddings 延后到有证据支持后再做
default card context 延后到 runtime evidence 证明后再做
Phase 3 baseline:
runtime adoption events 已存入 schema v9
adoption evidence 可通过 CLI inspect
card context default、card embeddings、evaluator APIs、research adapter ingestion 都有 readiness gates
research report validation 已存在,但不会自动 promotion
这已经足够进入新阶段,因为系统不再只是设计。它已经有 storage、lifecycle、 runtime context、MCP surfaces、card governance 和 measurement gates。
11. 哪些事情仍然故意没做
有些事情仍然故意不做。
Card-aware context 还不是默认行为。它仍然 opt-in,因为默认 runtime context 是 trust boundary。系统需要重复 runtime traces 证明它有收益,才能默认启用 cards。
Card embeddings 还不存在。当前 card retrieval 是 linked-evidence-first。这保留 citation semantics,并避免 stale card-vector maintenance。Card embeddings 需要 statement-match misses 和 rollback plan。
Research adapters 不会直接 ingest promoted knowledge。它们可以 validate input plans,未来也可以写 evidence 或 candidate insights,但 promotion 必须留在 memory lifecycle governance 下。
Evaluator APIs 不会 mutate knowledge state。它们可以 advise,但 gates 和 human-review requirements 仍然是权威边界。
系统现在也不宣称已经实现完全自主自进化。它有的是一个可治理的 self-improvement substrate:
evidence capture
candidate distillation
gate evaluation
reversible lifecycle
runtime adoption measurement
Autonomy 只能在有证据证明有效,并且 rollback 明确时逐步加入。
12. 后续设计原则
下一阶段应该用一个问题来判断每个改动:
这个改动是否能产生可测量的 runtime improvement,同时不削弱 evidence、 citation、audit 或 rollback 边界?
如果答案是否定的,这个改动就应该保持 advisory、opt-in 或 deferred。
这也是为什么下一步更适合先把 Phase 3 runtime surfaces 暴露给 MCP,并定义 adoption-recording protocol。Agent 需要直接记录某个 context item、card、 evaluator suggestion 或 research result 是否被 used、accepted、rejected、missed 或 rolled back。
只有这些 evidence 积累起来之后,系统才应该考虑更强默认行为:
card context default-on
card embeddings
evaluator APIs
research adapter apply paths
所以心智模型的推进顺序应该是:
定义 knowledge hierarchy
分离 evidence 和 knowledge
实现 lifecycle governance
暴露 runtime context
抽取 governed cards
保持 citations 和 audit
测量 runtime adoption
最后才增加 authority
这个顺序就是当前实现最重要的成果。
它把 memory 从被动 recall 变成了 governed learning loop。
它让 dao 留在 memory,shu 留在可复用方法,qi 留在工具层,而 evidence
成为所有 durable belief 的 justification substrate。
它也给下一阶段留下了明确任务:
从已经实现的 cognition architecture,推进到可测量、可回滚、证据驱动的 agent evolution。
Source Notes
本文基于仓库中的设计和实现清单:
docs/MIND-MODEL-DESIGN.mdAGENTS.mdCLAUDE.mdspecs/p12-mind-model-bootstrap.spec.mdspecs/p14-context-assembler.spec.mdspecs/p15-mcp-context.spec.mdspecs/p17-knowledge-lifecycle.spec.mdspecs/p18-knowledge-distill.spec.mdspecs/p20-promotion-gate-policy.spec.mdspecs/p31-knowledge-card-schema.spec.mdspecs/p32-knowledge-card-schema-v8.spec.mdspecs/p38-knowledge-card-gate.spec.mdspecs/p44-card-context-assembler.spec.mdspecs/p45-card-linked-evidence-retrieval.spec.mdspecs/p49-research-ingestion-policy.spec.mdspecs/p50-evaluator-promotion-policy.spec.mdspecs/p51-mind-model-closure-audit.spec.mdspecs/p52-phase-3-intake-roadmap.spec.mdspecs/p53-phase-3-candidate-evidence-audit.spec.mdspecs/p54-runtime-adoption-evidence.spec.mdspecs/p55-runtime-adoption-cli.spec.mdspecs/p56-card-context-default-gate.spec.mdspecs/p57-card-embedding-evidence-gate.spec.mdspecs/p58-evaluator-api-evidence-gate.spec.mdspecs/p59-research-adapter-ingestion-contract.spec.md
同时交叉核对了已存入 mempal 的 evidence:
drawer_mempal_p31_cc86ab94,source_file=mempal-p31-knowledge-card-schema-spec.mddrawer_mempal_p36_3770af58,source_file=mempal-p36-knowledge-card-backfill-report.mddrawer_mempal_p18_b6dac6fe,source_file=mempal-p18-knowledge-distill-memory.mddrawer_mempal_p30_7451b74c,source_file=mempal-p30-knowledge-card-storage-boundary.mddrawer_mempal_p53_82a98e1f,source_file=mempal-p53-phase-3-candidate-evidence-audit.mddrawer_mempal_p54_p59_866027ab,source_file=mempal-p54-p59-phase3-runtime.md
453 files · 436 not served here
- .claude/hooks/user-prompt-submit.sh (not served here)
- .claude/settings.json
- .github/workflows/ci.yml (not served here)
- .github/workflows/pages.yml (not served here)
- .github/workflows/release.yml (not served here)
- .gitignore (not served here)
- .mcp.json
- AGENTS.md
- ATTRIBUTION.md
- AUTOGINEER-RUNS.md
- CHANGELOG.md
- CLAUDE.md
- Cargo.lock (not served here)
- Cargo.toml (not served here)
- LICENSE
- README.md
- README_zh.md
- benchmarks/longmemeval_s_summary.md (not served here)
- books/README.md (not served here)
- books/zh-CN/book.toml (not served here)
- books/zh-CN/mermaid-init.js (not served here)
- books/zh-CN/mermaid.min.js (not served here)
- books/zh-CN/plans/00-preface.plan.md (not served here)
- books/zh-CN/plans/01-problem.plan.md (not served here)
- books/zh-CN/plans/02-principles.plan.md (not served here)
- books/zh-CN/plans/03-architecture.plan.md (not served here)
- books/zh-CN/plans/04-storage-search.plan.md (not served here)
- books/zh-CN/plans/05-mind-model.plan.md (not served here)
- books/zh-CN/plans/06-knowledge-governance.plan.md (not served here)
- books/zh-CN/plans/07-runtime-context.plan.md (not served here)
- books/zh-CN/plans/08-cowork.plan.md (not served here)
- books/zh-CN/plans/09-self-evolution.plan.md (not served here)
- books/zh-CN/plans/10-ops-and-usage.plan.md (not served here)
- books/zh-CN/plans/99-appendix-commands.plan.md (not served here)
- books/zh-CN/specs/00-preface.spec.md (not served here)
- books/zh-CN/specs/01-problem.spec.md (not served here)
- books/zh-CN/specs/02-principles.spec.md (not served here)
- books/zh-CN/specs/03-architecture.spec.md (not served here)
- books/zh-CN/specs/04-storage-search.spec.md (not served here)
- books/zh-CN/specs/05-mind-model.spec.md (not served here)
- books/zh-CN/specs/06-knowledge-governance.spec.md (not served here)
- books/zh-CN/specs/07-runtime-context.spec.md (not served here)
- books/zh-CN/specs/08-cowork.spec.md (not served here)
- books/zh-CN/specs/09-self-evolution.spec.md (not served here)
- books/zh-CN/specs/10-ops-and-usage.spec.md (not served here)
- books/zh-CN/specs/99-appendix-commands.spec.md (not served here)
- books/zh-CN/src/SUMMARY.md (not served here)
- books/zh-CN/src/appendix-commands.md (not served here)
- books/zh-CN/src/ch01-problem.md (not served here)
- books/zh-CN/src/ch02-principles.md (not served here)
- books/zh-CN/src/ch03-architecture.md (not served here)
- books/zh-CN/src/ch04-storage-search.md (not served here)
- books/zh-CN/src/ch05-mind-model.md (not served here)
- books/zh-CN/src/ch06-knowledge-governance.md (not served here)
- books/zh-CN/src/ch07-runtime-context.md (not served here)
- books/zh-CN/src/ch08-cowork.md (not served here)
- books/zh-CN/src/ch09-self-evolution.md (not served here)
- books/zh-CN/src/ch10-ops-and-usage.md (not served here)
- books/zh-CN/src/preface.md (not served here)
- crates/mempal-agent-memory/Cargo.toml (not served here)
- crates/mempal-agent-memory/src/anchor.rs (not served here)
- crates/mempal-agent-memory/src/lib.rs (not served here)
- crates/mempal-agent-memory/src/types.rs (not served here)
- crates/mempal-embed/Cargo.toml (not served here)
- crates/mempal-embed/src/api.rs (not served here)
- crates/mempal-embed/src/lib.rs (not served here)
- crates/mempal-embed/src/model2vec.rs (not served here)
- crates/mempal-embed/src/onnx.rs (not served here)
- crates/mempal-mcp-protocol/Cargo.toml (not served here)
- crates/mempal-mcp-protocol/src/lib.rs (not served here)
- crates/mempal-mcp-server/Cargo.toml (not served here)
- crates/mempal-mcp-server/src/lib.rs (not served here)
- crates/mempal-mcp-server/src/server.rs (not served here)
- crates/mempal-mcp-server/src/tools.rs (not served here)
- crates/mempal-runtime/Cargo.toml (not served here)
- crates/mempal-runtime/src/aaak/codec.rs (not served here)
- crates/mempal-runtime/src/aaak/mod.rs (not served here)
- crates/mempal-runtime/src/aaak/model.rs (not served here)
- crates/mempal-runtime/src/aaak/parse.rs (not served here)
- crates/mempal-runtime/src/aaak/signals.rs (not served here)
- crates/mempal-runtime/src/aaak/spec.rs (not served here)
- crates/mempal-runtime/src/adoption_analytics.rs (not served here)
- crates/mempal-runtime/src/brief.rs (not served here)
- crates/mempal-runtime/src/context.rs (not served here)
- crates/mempal-runtime/src/core/config.rs (not served here)
- crates/mempal-runtime/src/core/mod.rs (not served here)
- crates/mempal-runtime/src/core/phase3.rs (not served here)
- crates/mempal-runtime/src/core/protocol.rs (not served here)
- crates/mempal-runtime/src/core/utils.rs (not served here)
- crates/mempal-runtime/src/cowork/bus.rs (not served here)
- crates/mempal-runtime/src/cowork/claude.rs (not served here)
- crates/mempal-runtime/src/cowork/codex.rs (not served here)
- crates/mempal-runtime/src/cowork/inbox.rs (not served here)
- crates/mempal-runtime/src/cowork/mod.rs (not served here)
- crates/mempal-runtime/src/cowork/peek.rs (not served here)
- crates/mempal-runtime/src/cowork/receipts.rs (not served here)
- crates/mempal-runtime/src/doctor.rs (not served here)
- crates/mempal-runtime/src/embed/mod.rs (not served here)
- crates/mempal-runtime/src/factcheck/contradictions.rs (not served here)
- crates/mempal-runtime/src/factcheck/mod.rs (not served here)
- crates/mempal-runtime/src/factcheck/names.rs (not served here)
- crates/mempal-runtime/src/factcheck/relations.rs (not served here)
- crates/mempal-runtime/src/field_taxonomy.rs (not served here)
- crates/mempal-runtime/src/ingest/chunk.rs (not served here)
- crates/mempal-runtime/src/ingest/detect.rs (not served here)
- crates/mempal-runtime/src/ingest/diary.rs (not served here)
- crates/mempal-runtime/src/ingest/lock.rs (not served here)
- crates/mempal-runtime/src/ingest/mod.rs (not served here)
- crates/mempal-runtime/src/ingest/noise.rs (not served here)
- crates/mempal-runtime/src/ingest/normalize.rs (not served here)
- crates/mempal-runtime/src/ingest/reindex.rs (not served here)
- crates/mempal-runtime/src/knowledge_anchor.rs (not served here)
- crates/mempal-runtime/src/knowledge_card_backfill.rs (not served here)
- crates/mempal-runtime/src/knowledge_card_lifecycle.rs (not served here)
- crates/mempal-runtime/src/knowledge_card_retrieval.rs (not served here)
- crates/mempal-runtime/src/knowledge_distill.rs (not served here)
- crates/mempal-runtime/src/knowledge_gate.rs (not served here)
- crates/mempal-runtime/src/knowledge_lifecycle.rs (not served here)
- crates/mempal-runtime/src/lib.rs (not served here)
- crates/mempal-runtime/src/longmemeval.rs (not served here)
- crates/mempal-runtime/src/path_filter.rs (not served here)
- crates/mempal-runtime/src/projects.rs (not served here)
- crates/mempal-runtime/src/search/filter.rs (not served here)
- crates/mempal-runtime/src/search/mod.rs (not served here)
- crates/mempal-runtime/src/search/rerank.rs (not served here)
- crates/mempal-runtime/src/search/route.rs (not served here)
- crates/mempal-search-core/Cargo.toml (not served here)
- crates/mempal-search-core/src/lib.rs (not served here)
- crates/mempal-store-sqlite/Cargo.toml (not served here)
- crates/mempal-store-sqlite/src/lib.rs (not served here)
- docs/COWORK-RUNBOOK.md
- docs/MAINTENANCE-RUNBOOK.md
- docs/MIND-MODEL-DESIGN.md
- docs/MIND-MODEL-IMPLEMENTATION-ARTICLE.md
- docs/MIND-MODEL-IMPLEMENTATION-ARTICLE.zh-CN.md
- docs/aaak-dialect.md
- docs/plans/2026-04-08-p0-implementation.md (not served here)
- docs/plans/2026-04-09-p1-p4-implementation.md (not served here)
- docs/plans/2026-04-11-p5-implementation.md (not served here)
- docs/plans/2026-04-13-p6-implementation.md (not served here)
- docs/plans/2026-04-13-p7-implementation.md (not served here)
- docs/plans/2026-04-15-p8-implementation.md (not served here)
- docs/plans/2026-04-17-p9-implementation.md (not served here)
- docs/plans/2026-04-21-p12-implementation.md (not served here)
- docs/plans/2026-04-23-p10-explicit-tunnels-implementation.md (not served here)
- docs/plans/2026-04-23-p10-normalize-version-implementation.md (not served here)
- docs/plans/2026-04-23-p11-transcript-noise-strip-implementation.md (not served here)
- docs/plans/2026-04-23-p13a-implementation.md (not served here)
- docs/plans/2026-04-23-p13b-implementation.md (not served here)
- docs/plans/2026-04-24-p11-chunk-neighbors-implementation.md (not served here)
- docs/plans/2026-04-24-p11-diary-daily-rollup-implementation.md (not served here)
- docs/plans/2026-04-24-p14-context-assembler-implementation.md (not served here)
- docs/plans/2026-04-24-p15-mcp-context-implementation.md (not served here)
- docs/plans/2026-04-24-p16-context-skill-guidance-implementation.md (not served here)
- docs/plans/2026-04-24-p17-knowledge-lifecycle-implementation.md (not served here)
- docs/plans/2026-04-24-p18-knowledge-distill-implementation.md (not served here)
- docs/plans/2026-04-24-p19-lifecycle-ref-validation-implementation.md (not served here)
- docs/plans/2026-04-25-p20-promotion-gate-policy-implementation.md (not served here)
- docs/plans/2026-04-25-p21-mcp-knowledge-gate-implementation.md (not served here)
- docs/plans/2026-04-25-p22-mcp-knowledge-distill-implementation.md (not served here)
- docs/plans/2026-04-26-p23-mcp-knowledge-lifecycle-implementation.md (not served here)
- docs/plans/2026-04-26-p24-anchor-publication-implementation.md (not served here)
- docs/plans/2026-04-26-p25-mcp-anchor-publication-implementation.md (not served here)
- docs/plans/2026-04-26-p26-dao-tian-runtime-budget-implementation.md (not served here)
- docs/plans/2026-04-26-p27-knowledge-policy-surface-implementation.md (not served here)
- docs/plans/2026-04-26-p28-field-taxonomy-surface-implementation.md (not served here)
- docs/plans/2026-04-26-p29-wake-up-context-boundary-implementation.md (not served here)
- docs/plans/2026-04-27-p30-knowledge-card-storage-boundary-implementation.md (not served here)
- docs/plans/2026-04-27-p31-knowledge-card-schema-spec.md (not served here)
- docs/plans/2026-04-27-p32-knowledge-card-schema-v8-implementation.md (not served here)
- docs/plans/2026-04-27-p33-knowledge-card-core-api-implementation.md (not served here)
- docs/plans/2026-04-27-p34-knowledge-card-cli-implementation.md (not served here)
- docs/plans/2026-04-27-p35-knowledge-card-mcp-read-implementation.md (not served here)
- docs/plans/2026-04-27-p36-knowledge-card-backfill-report-implementation.md (not served here)
- docs/plans/2026-04-27-p37-knowledge-card-backfill-apply-implementation.md (not served here)
- docs/plans/2026-04-27-p38-p42-knowledge-card-runtime-implementation.md (not served here)
- docs/plans/2026-04-28-p43-knowledge-card-retrieval-contract.md (not served here)
- docs/plans/2026-04-28-p44-card-context-assembler.md (not served here)
- docs/plans/2026-04-28-p45-card-linked-evidence-retrieval.md (not served here)
- docs/plans/2026-04-28-p46-card-context-default-policy.md (not served here)
- docs/plans/2026-04-28-p47-card-embedding-policy.md (not served here)
- docs/plans/2026-04-29-p48-card-audit-policy.md (not served here)
- docs/plans/2026-04-29-p49-research-ingestion-policy.md (not served here)
- docs/plans/2026-04-29-p50-evaluator-promotion-policy.md (not served here)
- docs/plans/2026-04-29-p51-mind-model-closure-audit.md (not served here)
- docs/plans/2026-05-01-p52-phase-3-intake-roadmap.md (not served here)
- docs/plans/2026-05-02-p53-phase-3-candidate-evidence-audit.md (not served here)
- docs/plans/2026-05-02-p54-runtime-adoption-evidence.md (not served here)
- docs/plans/2026-05-02-p55-runtime-adoption-cli.md (not served here)
- docs/plans/2026-05-02-p56-card-context-default-gate.md (not served here)
- docs/plans/2026-05-02-p57-card-embedding-evidence-gate.md (not served here)
- docs/plans/2026-05-02-p58-evaluator-api-evidence-gate.md (not served here)
- docs/plans/2026-05-02-p59-research-adapter-ingestion-contract.md (not served here)
- docs/plans/2026-05-05-p60-mcp-phase3-runtime-surface.md (not served here)
- docs/plans/2026-05-10-p61-runtime-adoption-recording-protocol.md (not served here)
- docs/plans/2026-05-12-p62-runtime-adoption-cli-guidance.md (not served here)
- docs/plans/2026-05-12-p63-runtime-adoption-record-helper.md (not served here)
- docs/plans/2026-05-13-p64-runtime-adoption-record-quality-policy.md (not served here)
- docs/plans/2026-05-13-p65-runtime-adoption-review-report.md (not served here)
- docs/plans/2026-05-13-p66-card-context-default-readiness.md (not served here)
- docs/plans/2026-05-13-p67-research-adapter-ingest.md (not served here)
- docs/plans/2026-05-13-p68-mcp-research-ingest-plan.md (not served here)
- docs/plans/2026-05-13-p69-runtime-adoption-checked-record.md (not served here)
- docs/plans/2026-05-13-p70-self-evolution-completion-audit.md (not served here)
- docs/plans/2026-05-13-p71-self-evolution-loop-replay.md (not served here)
- docs/plans/2026-05-13-p72-runtime-adoption-capture-helper.md (not served here)
- docs/plans/2026-05-13-p73-evaluator-advisory-api.md (not served here)
- docs/plans/2026-05-13-p74-card-context-default-proposal.md (not served here)
- docs/plans/2026-05-13-p75-self-evolution-completion-audit.md (not served here)
- docs/plans/2026-05-13-p76-spec-completeness-invariant.md (not served here)
- docs/plans/2026-05-13-p77-live-adoption-instrumentation-boundary.md (not served here)
- docs/plans/2026-05-13-p78-card-context-default-runtime-flag.md (not served here)
- docs/plans/2026-05-13-p79-rollback-executor-policy.md (not served here)
- docs/plans/2026-05-13-p80-autonomous-promotion-boundary-audit.md (not served here)
- docs/plans/2026-05-13-p81-self-evolution-completion-audit.md (not served here)
- docs/plans/2026-05-25-p82-opt-in-runtime-instrumentation-wrapper.md (not served here)
- docs/plans/2026-05-25-p83-cognitive-brief.md (not served here)
- docs/plans/2026-05-25-p84-multi-agent-cowork-bus.md (not served here)
- docs/plans/2026-05-25-p85-mcp-multi-agent-cowork-bus.md (not served here)
- docs/plans/2026-05-25-p86-tmux-cowork-transport.md (not served here)
- docs/plans/2026-05-25-p87-cowork-bus-event-log.md (not served here)
- docs/plans/2026-05-25-p88-cowork-delivery-ack-status.md (not served here)
- docs/plans/2026-05-25-p89-cowork-agent-presence.md (not served here)
- docs/plans/2026-05-25-p90-cowork-threads-channels.md (not served here)
- docs/plans/2026-05-25-p91-tmux-live-peek.md (not served here)
- docs/plans/2026-05-26-p100-guided-maintenance-run.md (not served here)
- docs/plans/2026-05-26-p101-cowork-session-close-capture.md (not served here)
- docs/plans/2026-05-26-p102-mcp-cognitive-brief.md (not served here)
- docs/plans/2026-05-26-p103-adoption-analytics.md (not served here)
- docs/plans/2026-05-26-p104-release-readiness-checklist.md (not served here)
- docs/plans/2026-05-26-p92-multi-agent-runbook.md (not served here)
- docs/plans/2026-05-26-p93-cowork-doctor.md (not served here)
- docs/plans/2026-05-26-p94-cowork-team-session.md (not served here)
- docs/plans/2026-05-26-p95-cowork-handoff-summary.md (not served here)
- docs/plans/2026-05-26-p96-cowork-memory-capture.md (not served here)
- docs/plans/2026-05-26-p97-maintenance-runbook.md (not served here)
- docs/plans/2026-05-26-p98-release-install-doctor.md (not served here)
- docs/plans/2026-05-26-p99-mcp-runtime-doctor.md (not served here)
- docs/plans/2026-05-28-p105-mempal-book-zh-cn.md (not served here)
- docs/plans/2026-06-04-p106-context-distill-signal.md (not served here)
- docs/plans/2026-06-09-p107-ingest-entrypoint-discoverability.md (not served here)
- docs/plans/2026-06-17-p108-peek-cwd-cross-project.md (not served here)
- docs/plans/2026-06-22-p109-cross-project-resume.md (not served here)
- docs/plans/2026-07-02-p110-public-issue-compatibility.md (not served here)
- docs/plans/2026-07-06-p111-project-ignore-rules.md (not served here)
- docs/plans/2026-07-07-p112-session-peek.md (not served here)
- docs/plans/2026-07-08-p113-public-workspace-crates.md (not served here)
- docs/plans/2026-07-09-p114-coarse-workspace-split.md (not served here)
- docs/plans/2026-07-26-p115-codex-preamble-noise-strip.md (not served here)
- docs/plans/2026-07-26-p116-cowork-pair-delivery-receipts.md (not served here)
- docs/plans/2026-08-18-p117-reindex-atomic-replace.md (not served here)
- docs/site/en/index.html (not served here)
- docs/site/index.html (not served here)
- docs/site/ja/index.html (not served here)
- docs/specs/2026-04-08-mempal-design.md (not served here)
- docs/specs/2026-04-13-cowork-peek-and-decide.md (not served here)
- docs/specs/2026-04-13-p7-search-structured-signals.md (not served here)
- docs/specs/2026-04-14-p8-cowork-inbox-push.md (not served here)
- docs/usage.md
- hooks/mempal_precompact_hook.sh (not served here)
- hooks/mempal_save_hook.sh (not served here)
- specs/p0-core-scaffold.spec.md (not served here)
- specs/p0-embed-trait.spec.md (not served here)
- specs/p0-ingest.spec.md (not served here)
- specs/p0-search-cli.spec.md (not served here)
- specs/p1-routing-citation.spec.md (not served here)
- specs/p10-explicit-tunnels.spec.md (not served here)
- specs/p10-normalize-version.spec.md (not served here)
- specs/p100-guided-maintenance-run.spec.md (not served here)
- specs/p101-cowork-session-close-capture.spec.md (not served here)
- specs/p102-mcp-cognitive-brief.spec.md (not served here)
- specs/p103-adoption-analytics.spec.md (not served here)
- specs/p104-release-readiness-checklist.spec.md (not served here)
- specs/p105-mempal-book-zh-cn.spec.md (not served here)
- specs/p106-context-distill-signal.spec.md (not served here)
- specs/p107-ingest-entrypoint-discoverability.spec.md (not served here)
- specs/p108-peek-cwd-cross-project.spec.md (not served here)
- specs/p109-cross-project-resume.spec.md (not served here)
- specs/p11-chunk-neighbors.spec.md (not served here)
- specs/p11-diary-daily-rollup.spec.md (not served here)
- specs/p11-transcript-noise-strip.spec.md (not served here)
- specs/p110-public-issue-compatibility.spec.md (not served here)
- specs/p111-project-ignore-rules.spec.md (not served here)
- specs/p112-session-peek.spec.md (not served here)
- specs/p113-public-workspace-crates.spec.md (not served here)
- specs/p114-coarse-workspace-split.spec.md (not served here)
- specs/p115-codex-preamble-noise-strip.spec.md (not served here)
- specs/p116-cowork-pair-delivery-receipts.spec.md (not served here)
- specs/p117-reindex-atomic-replace.spec.md (not served here)
- specs/p12-mind-model-bootstrap.spec.md (not served here)
- specs/p13-ingest-identity.spec.md (not served here)
- specs/p13-wake-up-statement.spec.md (not served here)
- specs/p14-context-assembler.spec.md (not served here)
- specs/p15-mcp-context.spec.md (not served here)
- specs/p16-context-skill-guidance.spec.md (not served here)
- specs/p17-knowledge-lifecycle.spec.md (not served here)
- specs/p18-knowledge-distill.spec.md (not served here)
- specs/p19-lifecycle-ref-validation.spec.md (not served here)
- specs/p2-mcp.spec.md (not served here)
- specs/p20-promotion-gate-policy.spec.md (not served here)
- specs/p21-mcp-knowledge-gate.spec.md (not served here)
- specs/p22-mcp-knowledge-distill.spec.md (not served here)
- specs/p23-mcp-knowledge-lifecycle.spec.md (not served here)
- specs/p24-anchor-publication.spec.md (not served here)
- specs/p25-mcp-anchor-publication.spec.md (not served here)
- specs/p26-dao-tian-runtime-budget.spec.md (not served here)
- specs/p27-knowledge-policy-surface.spec.md (not served here)
- specs/p28-field-taxonomy-surface.spec.md (not served here)
- specs/p29-wake-up-context-boundary.spec.md (not served here)
- specs/p3-aaak.spec.md (not served here)
- specs/p30-knowledge-card-storage-boundary.spec.md (not served here)
- specs/p31-knowledge-card-schema.spec.md (not served here)
- specs/p32-knowledge-card-schema-v8.spec.md (not served here)
- specs/p33-knowledge-card-core-api.spec.md (not served here)
- specs/p34-knowledge-card-cli.spec.md (not served here)
- specs/p35-knowledge-card-mcp-read.spec.md (not served here)
- specs/p36-knowledge-card-backfill-report.spec.md (not served here)
- specs/p37-knowledge-card-backfill-apply.spec.md (not served here)
- specs/p38-knowledge-card-gate.spec.md (not served here)
- specs/p39-knowledge-card-lifecycle-cli.spec.md (not served here)
- specs/p4-rest-api.spec.md (not served here)
- specs/p40-mcp-knowledge-card-lifecycle.spec.md (not served here)
- specs/p41-knowledge-card-runtime-boundary.spec.md (not served here)
- specs/p42-mind-model-completion-audit.spec.md (not served here)
- specs/p43-knowledge-card-retrieval-contract.spec.md (not served here)
- specs/p44-card-context-assembler.spec.md (not served here)
- specs/p45-card-linked-evidence-retrieval.spec.md (not served here)
- specs/p46-card-context-default-policy.spec.md (not served here)
- specs/p47-card-embedding-policy.spec.md (not served here)
- specs/p48-card-audit-policy.spec.md (not served here)
- specs/p49-research-ingestion-policy.spec.md (not served here)
- specs/p5-agent-diary.spec.md (not served here)
- specs/p5-format-support.spec.md (not served here)
- specs/p5-kg-timeline-stats.spec.md (not served here)
- specs/p5-semantic-dedup.spec.md (not served here)
- specs/p5-wake-up-importance.spec.md (not served here)
- specs/p50-evaluator-promotion-policy.spec.md (not served here)
- specs/p51-mind-model-closure-audit.spec.md (not served here)
- specs/p52-phase-3-intake-roadmap.spec.md (not served here)
- specs/p53-phase-3-candidate-evidence-audit.spec.md (not served here)
- specs/p54-runtime-adoption-evidence.spec.md (not served here)
- specs/p55-runtime-adoption-cli.spec.md (not served here)
- specs/p56-card-context-default-gate.spec.md (not served here)
- specs/p57-card-embedding-evidence-gate.spec.md (not served here)
- specs/p58-evaluator-api-evidence-gate.spec.md (not served here)
- specs/p59-research-adapter-ingestion-contract.spec.md (not served here)
- specs/p6-cowork-peek-and-decide.spec.md (not served here)
- specs/p60-mcp-phase3-runtime-surface.spec.md (not served here)
- specs/p61-runtime-adoption-recording-protocol.spec.md (not served here)
- specs/p62-runtime-adoption-cli-guidance.spec.md (not served here)
- specs/p63-runtime-adoption-record-helper.spec.md (not served here)
- specs/p64-runtime-adoption-record-quality-policy.spec.md (not served here)
- specs/p65-runtime-adoption-review-report.spec.md (not served here)
- specs/p66-card-context-default-readiness.spec.md (not served here)
- specs/p67-research-adapter-ingest.spec.md (not served here)
- specs/p68-mcp-research-ingest-plan.spec.md (not served here)
- specs/p69-runtime-adoption-checked-record.spec.md (not served here)
- specs/p7-search-structured-signals.spec.md (not served here)
- specs/p70-self-evolution-completion-audit.spec.md (not served here)
- specs/p71-self-evolution-loop-replay.spec.md (not served here)
- specs/p72-runtime-adoption-capture-helper.spec.md (not served here)
- specs/p73-evaluator-advisory-api.spec.md (not served here)
- specs/p74-card-context-default-proposal.spec.md (not served here)
- specs/p75-self-evolution-completion-audit.spec.md (not served here)
- specs/p76-spec-completeness-invariant.spec.md (not served here)
- specs/p77-live-adoption-instrumentation-boundary.spec.md (not served here)
- specs/p78-card-context-default-runtime-flag.spec.md (not served here)
- specs/p79-rollback-executor-policy.spec.md (not served here)
- specs/p8-cowork-inbox-push.spec.md (not served here)
- specs/p80-autonomous-promotion-boundary-audit.spec.md (not served here)
- specs/p81-self-evolution-completion-audit.spec.md (not served here)
- specs/p82-opt-in-runtime-instrumentation-wrapper.spec.md (not served here)
- specs/p83-cognitive-brief.spec.md (not served here)
- specs/p84-multi-agent-cowork-bus.spec.md (not served here)
- specs/p85-mcp-multi-agent-cowork-bus.spec.md (not served here)
- specs/p86-tmux-cowork-transport.spec.md (not served here)
- specs/p87-cowork-bus-event-log.spec.md (not served here)
- specs/p88-cowork-delivery-ack-status.spec.md (not served here)
- specs/p89-cowork-agent-presence.spec.md (not served here)
- specs/p9-fact-checker.spec.md (not served here)
- specs/p9-ingest-lock.spec.md (not served here)
- specs/p90-cowork-threads-channels.spec.md (not served here)
- specs/p91-tmux-live-peek.spec.md (not served here)
- specs/p92-multi-agent-runbook.spec.md (not served here)
- specs/p93-cowork-doctor.spec.md (not served here)
- specs/p94-cowork-team-session.spec.md (not served here)
- specs/p95-cowork-handoff-summary.spec.md (not served here)
- specs/p96-cowork-memory-capture.spec.md (not served here)
- specs/p97-maintenance-runbook.spec.md (not served here)
- specs/p98-release-install-doctor.spec.md (not served here)
- specs/p99-mcp-runtime-doctor.spec.md (not served here)
- specs/project.spec.md (not served here)
- src/aaak/mod.rs (not served here)
- src/adoption_analytics.rs (not served here)
- src/api/handlers.rs (not served here)
- src/api/mod.rs (not served here)
- src/api/state.rs (not served here)
- src/brief.rs (not served here)
- src/context.rs (not served here)
- src/core/db.rs (not served here)
- src/core/mod.rs (not served here)
- src/core/protocol.rs (not served here)
- src/cowork/mod.rs (not served here)
- src/doctor.rs (not served here)
- src/embed/mod.rs (not served here)
- src/factcheck/mod.rs (not served here)
- src/field_taxonomy.rs (not served here)
- src/ingest/mod.rs (not served here)
- src/knowledge_anchor.rs (not served here)
- src/knowledge_card_backfill.rs (not served here)
- src/knowledge_card_lifecycle.rs (not served here)
- src/knowledge_card_retrieval.rs (not served here)
- src/knowledge_distill.rs (not served here)
- src/knowledge_gate.rs (not served here)
- src/knowledge_lifecycle.rs (not served here)
- src/lib.rs (not served here)
- src/longmemeval.rs (not served here)
- src/main.rs (not served here)
- src/mcp/mod.rs (not served here)
- src/path_filter.rs (not served here)
- src/projects.rs (not served here)
- src/search/mod.rs (not served here)
- tests/cognitive_brief.rs (not served here)
- tests/context_assembler.rs (not served here)
- tests/cowork_bus.rs (not served here)
- tests/cowork_inbox.rs (not served here)
- tests/cowork_peek.rs (not served here)
- tests/cowork_receipts.rs (not served here)
- tests/diary_rollup.rs (not served here)
- tests/fact_check.rs (not served here)
- tests/fixtures/cowork/claude/session.jsonl (not served here)
- tests/fixtures/cowork/claude/session_with_tools.jsonl (not served here)
- tests/fixtures/cowork/claude_since/session.jsonl (not served here)
- tests/fixtures/cowork/codex/2026/04/12/rollout-2026-04-12T12-00-00-otherproject.jsonl (not served here)
- tests/fixtures/cowork/codex/2026/04/13/rollout-2026-04-13T12-00-00-fake.jsonl (not served here)
- tests/ingest_lock.rs (not served here)
- tests/knowledge_card_backfill.rs (not served here)
- tests/knowledge_card_cli.rs (not served here)
- tests/knowledge_card_core.rs (not served here)
- tests/knowledge_card_retrieval.rs (not served here)
- tests/knowledge_card_schema.rs (not served here)
- tests/knowledge_lifecycle.rs (not served here)
- tests/mind_model_bootstrap.rs (not served here)
- tests/noise_strip.rs (not served here)
- tests/normalize_version.rs (not served here)
- tests/ops_runtime.rs (not served here)
- tests/phase3_runtime.rs (not served here)
- tests/phase3_self_evolution_replay.rs (not served here)
- tests/projects_resume.rs (not served here)
- tests/reindex_safety.rs (not served here)
- tests/search_neighbors.rs (not served here)
- tests/tunnels_explicit.rs (not served here)
- tests/workspace_crates.rs (not served here)
Advantages
- Hybrid search combines BM25 and vector semantic search for improved retrieval. evidence: README.md
- Self-describing protocol allows agents to learn interaction rules automatically. evidence: README.md
- Per-source advisory locking prevents race conditions during concurrent agent ingestion. evidence: README.md
Limitations
- Codex integration requires a specific feature flag to be enabled. evidence: README.md
- Bidirectional cowork push is restricted to identified Claude and Codex agents. evidence: README.md
- Advisory locking for concurrent ingest is currently limited to Unix systems. evidence: README.md
Security
Risk level: Low risk
This brain runs
Hooks (2)
.claude/settings.json(PostToolUse):if echo "$TOOL_INPUT" | grep -q 'git commit'; then echo 'MEMPAL REMINDER: You just committed code. Call mempal_ingest to save the decision context (what was built, why, what was considered). Rule 4: SAVE AFTER DECISIONS.'; fi.claude/settings.json(UserPromptSubmit):bash .claude/hooks/user-prompt-submit.sh
MCP servers (1)
.mcp.json(mempal):mempal serve --mcp
Agent instructions that run or fetch something (22)
AGENTS.md:Rust 实现的 coding agent 项目记忆工具。单二进制,`cargo install mempal`,10 秒内带出处找回历史决策。AGENTS.md:| Spec | 状态 | 范围 | |------|------|------| | `specs/p0-core-scaffold.spec.md` | 完成 | workspace 骨架 + SQLite schema | | `specs/p0-embed-trait.spec.md` | 完成 | Embedder trait(model2vec 默认 + ort 可选) | | `specs/p0-ingest.spec.md` | 完成 | 导入管道(格式检测/归一化/分块/存储) | | `specs/p0-search-cli.spec.md` …[+7715 chars]AGENTS.md:`docs/plans/2026-05-26-p98-release-install-doctor.md` — P98 release install doctor(已完成)AGENTS.md:`docs/plans/2026-05-26-p100-guided-maintenance-run.md` — P100 guided maintenance run(已完成)AGENTS.md:`docs/plans/2026-08-18-p117-reindex-atomic-replace.md` — P117 atomic reindex replace + dry-run feasibility report(已完成)AGENTS.md:| 工具 | 作用 | |------|------| | `mempal_status` | 状态 + 协议 + AAAK spec | | `mempal_doctor` | release/install + MCP runtime diagnostics:报告 binary/PATH/schema 兼容性和 required MCP tool/action 能力清单(P99) | | `mempal_search` | 混合检索(BM25 + 向量 + RRF + tunnel hints)+ AAAK 结构化 signals(P7) | | `mempa …[+2688 chars]AGENTS.md:Cargo workspace,root package 仍是 `mempal`,并继续提供 `cargo install mempal` 的 CLI binary。P113/P114 已把可复用边界拆成公开 workspace crates;root crate 现在主要负责 CLI、REST entrypoint、package metadata 和 legacy facade paths。AGENTS.md:**`mempal cowork-install-hooks` 写两件 Claude 侧制品**:`.claude/hooks/user-prompt-submit.sh`(脚本)+ `.claude/settings.json` 下的 `hooks.UserPromptSubmit` 条目(注册)。Claude Code 不按文件名约定自动发现脚本,两件**都必须有**hook 才会 fire。`install-hooks` 已自动处理 + 自愈 stale 条目;请勿手工移除其中任一。AGENTS.md:**Codex 侧依赖 `codex_hooks` feature flag**:shipped `codex-cli ≤ 0.120.0` 该 flag 处于 "under development" 且默认 `false`,此时 Codex runtime 完全忽略 `~/.codex/hooks.json`。`install-hooks` 检测到会打印 warning + 激活命令 `codex features enable codex_hooks`。AGENTS.md:**Codex TUI 进程启动时一次性缓存 config**:改完 `config.toml` 或 `hooks.json`(含 feature flag / install-hooks)后,必须完全退出并重启 Codex TUI;已在跑的进程拿不到新配置。AGENTS.md:**Claude Code 的 MCP server 是 session startup spawn**:`cargo install` 升级 mempal binary 后,Claude Code 还在用旧 MCP server 进程,**不认识新加的工具**(如 `mempal_cowork_push`)。升级后重启 Claude Code,MCP server 会 respawn 到新 binary。CLAUDE.md:Rust 实现的 coding agent 项目记忆工具。单二进制,`cargo install mempal`,10 秒内带出处找回历史决策。CLAUDE.md:| Spec | 状态 | 范围 | |------|------|------| | `specs/p0-core-scaffold.spec.md` | 完成 | workspace 骨架 + SQLite schema | | `specs/p0-embed-trait.spec.md` | 完成 | Embedder trait(model2vec 默认 + ort 可选) | | `specs/p0-ingest.spec.md` | 完成 | 导入管道(格式检测/归一化/分块/存储) | | `specs/p0-search-cli.spec.md` …[+7715 chars]CLAUDE.md:`docs/plans/2026-05-26-p98-release-install-doctor.md` — P98 release install doctor(已完成)CLAUDE.md:`docs/plans/2026-05-26-p100-guided-maintenance-run.md` — P100 guided maintenance run(已完成)CLAUDE.md:`docs/plans/2026-08-18-p117-reindex-atomic-replace.md` — P117 atomic reindex replace + dry-run feasibility report(已完成)CLAUDE.md:| 工具 | 作用 | |------|------| | `mempal_status` | 状态 + 协议 + AAAK spec | | `mempal_doctor` | release/install + MCP runtime diagnostics:报告 binary/PATH/schema 兼容性和 required MCP tool/action 能力清单(P99) | | `mempal_search` | 混合检索(BM25 + 向量 + RRF + tunnel hints)+ AAAK 结构化 signals(P7) | | `mempa …[+2688 chars]CLAUDE.md:Cargo workspace,root package 仍是 `mempal`,并继续提供 `cargo install mempal` 的 CLI binary。P113/P114 已把可复用边界拆成公开 workspace crates;root crate 现在主要负责 CLI、REST entrypoint、package metadata 和 legacy facade paths。CLAUDE.md:**`mempal cowork-install-hooks` 写两件 Claude 侧制品**:`.claude/hooks/user-prompt-submit.sh`(脚本)+ `.claude/settings.json` 下的 `hooks.UserPromptSubmit` 条目(注册)。Claude Code 不按文件名约定自动发现脚本,两件**都必须有**hook 才会 fire。`install-hooks` 已自动处理 + 自愈 stale 条目;请勿手工移除其中任一。CLAUDE.md:**Codex 侧依赖 `codex_hooks` feature flag**:shipped `codex-cli ≤ 0.120.0` 该 flag 处于 "under development" 且默认 `false`,此时 Codex runtime 完全忽略 `~/.codex/hooks.json`。`install-hooks` 检测到会打印 warning + 激活命令 `codex features enable codex_hooks`。CLAUDE.md:**Codex TUI 进程启动时一次性缓存 config**:改完 `config.toml` 或 `hooks.json`(含 feature flag / install-hooks)后,必须完全退出并重启 Codex TUI;已在跑的进程拿不到新配置。CLAUDE.md:**Claude Code 的 MCP server 是 session startup spawn**:`cargo install` 升级 mempal binary 后,Claude Code 还在用旧 MCP server 进程,**不认识新加的工具**(如 `mempal_cowork_push`)。升级后重启 Claude Code,MCP server 会 respawn 到新 binary。
- Low risk: scripts-present (.claude/hooks/user-prompt-submit.sh)
Attribution and licence
- First committed by
ZhangHanDong
- Publisher
- Alex (Community)
- Source
- ZhangHanDong/mempal · commit 9ed705d
- Licence
- MIT License (files served here)
- ATTRIBUTION.md (written by AutoGineer)
- Synced
- Contributors
- ZhangHanDong, LiuZiqi-dual, TigerInYourDream, jolestar, ryushione
AutoGineer is not affiliated with or endorsed by ZhangHanDong.
Install from source
436 of its files are not served here, so there is no ZIP. ZhangHanDong/mempal at 9ed705d