codex-supermemory

codex-supermemory

Persistent memory for OpenAI Codex CLI — powered by Supermemory

Codex forgets every session. codex-supermemory wires Supermemory into Codex CLI's hooks system so your coding agent remembers your stack, preferences, prior decisions, and the lessons learned across every project — automatically.

Features

• 🧠 Automatic recall — relevant memories are injected for substantive prompts via the UserPromptSubmit hook, with visible recall counts and a 3-second network cap.

• 🔎 Hosted MCP tools — deeper search and explicit memory operations use mcp.supermemory.ai through the same credentials as the hooks.

• 💾 Automatic capture — completed turns are saved in the background via the Stop hook.

• 🏷️ Shared Agents scoping — Codex, Claude Code, and OpenCode use one collision-safe repository container.

• 🏷️ Personal + project routing — sm_scope metadata keeps automatic/personal memories distinguishable from explicit project knowledge in the shared container.

• Entity-aware extraction - the shared container uses one coding-agent context covering durable preferences and project/codebase facts.

• 🔒 Privacy-aware — anything wrapped in <private>...</private> is redacted before being sent to Supermemory.

• ⚡ Zero-config install — one command sets up ~/.codex/config.toml and ~/.codex/hooks.json for you.

• 🪶 No runtime deps in hooks — the hook scripts are pre-bundled with esbuild for fast cold starts.

• 🔧 Status and index skills — $supermemory-status checks authentication and connectivity; $supermemory-index explores the repo and saves focused memories via MCP add_memory. Memory operations still come from MCP, not command skills.

• ◪ Persistent CLI mark — compatible Codex terminals keep a quiet Supermemory badge at the bottom of the TUI, while hook notices report live recall and save activity.

Quick start

1. Install the hooks:

   Copy

   npx codex-supermemory install

2. Start Codex CLI. On your first prompt, a browser window will open to authenticate with Supermemory automatically.

   Alternatively, set export SUPERMEMORY_CODEX_API_KEY="sm_..." in your shell profile.

3. That's it — memory is active.

How it works

Codex CLI supports hooks and MCP servers. codex-supermemory registers four hooks:

Prompt recall and automatic capture call the Supermemory API directly. Deeper model-initiated search, add, list, and forget operations go through the hosted MCP server. Pointing REST at a self-hosted server does not redirect or disable MCP; see self-hosted REST and MCP routing before using a local key.

The installer:

• Registers the supermemory MCP server in ~/.codex/config.toml

• Registers the hooks in ~/.codex/hooks.json

• Copies pre-bundled hook scripts to ~/.codex/supermemory/

• Installs the supermemory-status and supermemory-index skills to ~/.codex/skills/

• Installs a static custom TUI badge to ~/.codex/pets/supermemory/

The installer selects the badge only when no Codex pet preference already exists. Terminals without a supported inline-image protocol may not render it; recall and capture continue to work. Use Codex's /pet picker to disable or change the persistent badge.

The hooks are tolerant: if Supermemory is unreachable, the API key is missing, or anything else fails, they exit cleanly without breaking your Codex session.

Shared Agents containers

Codex, Claude Code, and OpenCode use one container for a repository:

• repo_<project-name>__<remote-hash> stores automatic capture and every explicit save.

• sm_scope metadata preserves optional personal/project filtering.

The hash comes from the normalized Git remote, so clones share memory while same-named repositories do not collide. Repositories without a remote fall back to a local path identity. Codex also reads the previous user_project_*, repo_<project-name>, codex_user_*, codex_project_*, claudecode_project_*, opencode_user_*, and opencode_project_* containers, so existing memories remain searchable without duplicating or migrating them. Set SUPERMEMORY_ISOLATE_WORKTREES=true to use the worktree path instead of the remote identity.

Explicit projectContainerTag/repoContainerTag overrides remain the canonical write destination. Older user/personal overrides remain in the legacy read set.

Configuration

API v5 and existing installations

Version 1.0.20 bundles the official [email protected] SDK for document capture, profiles, search, and the status connectivity probe. Existing container-tag strings become namespace paths without renaming stored data. Config keys, credential files, session document IDs, capture cursors, legacy namespace reads, and recall approvals stay in place; no historical backfill or namespace move runs. Re-run npx codex-supermemory@latest install after upgrading to refresh the copied standalone hooks and status script. Updating npm dependencies alone doesn't update scripts already installed in ~/.codex/supermemory.

The hosted root https://api.supermemory.ai defaults to v5, including equivalent host casing and default-port spellings. Custom REST URLs, including URL prefixes, default to the bundled official SDK4 compatibility path so servers older than 0.0.9 remain usable. After upgrading a custom server, set SUPERMEMORY_API_VERSION=v5 or apiVersion: "v5" in the existing config. Select legacy explicitly to use v3/v4 where the server still supports them. Invalid version selections fail the operation; a failed v5 request never switches API versions or sends data to another host. REST URL selection remains independent of the hosted MCP and browser-auth endpoints, including their existing env fields.

The v5 profile has no query; prompt recall runs profile and memory search in parallel. Separated search explicitly uses the guide's legacy defaults (memories, 0.6 threshold, 10 candidates, no reranking/query rewriting), then preserves the local 0.55 floor, cross-namespace deduplication, and top-five/configured cap. The SDK search helper retains its existing hybrid mode and configured threshold and limit. Server-side ranking and historical-data availability aren't verified by local contract tests. Capture keeps its three-second budget and zero SDK retries, uses POST append/diff with the existing session ID, and advances the cursor only after valid acceptance. Dynamic processing is asynchronous; acceptance doesn't prove extraction completion or exactly-once billing after a lost response.

The retained dist/services/client.js library exports still have their original 30-second local wait and two-retry SDK cap for calls without hook options. SDK5's retryable statuses/backoff differ from SDK4, including no automatic 409 retry. Installed automatic hooks don't use that unbounded policy: a conflict keeps the capture cursor for a later attempt without an immediate retry. Explicit MCP operations still use their independent transport rather than these SDK helpers.

Display-name updates (/v3/container-tags/...) and the latent exact-content forget helper (/v4/memories) remain legacy-only ancillary operations against the configured REST server. Account details use /session, falling back to the deprecated /v3/session on older servers. They have no mechanical v5 equivalent. A v5-only server may leave names/account details unavailable or return an explicit forget error; the plugin does not substitute semantic deletion. Interactive MCP tools keep their independent protocol, endpoints, and approvals.

Self-hosted REST and MCP routing

The plugin has separate REST and MCP transports. A local REST server with MCP enabled is a hybrid setup, not an all-local memory integration: automatic recall/capture use REST, while explicit search, add, list, and forget requests use MCP. The self-hosting feature matrix lists Supermemory MCP as platform-only. A REST URL override neither provides a local MCP implementation nor changes the proxy's destination.

REST URLs must be absolute HTTP(S) URLs. SUPERMEMORY_API_VERSION/apiVersion selects the REST API version, not its host or the MCP protocol. There is no mcpUrl setting in supermemory.json; use the MCP process environment instead.

Both transports resolve their API key from nonempty SUPERMEMORY_CODEX_API_KEY in their own process, then apiKey in ~/.codex/supermemory.json, then the saved credential's apiKey. The MCP proxy sends that key as a Bearer credential to its selected MCP endpoint. Use a hosted-platform key for hosted MCP; do not assume a key generated by the local server is valid there. Leaving MCP enabled with a local key still sends that credential to the selected MCP endpoint.

Configure local REST

Merge these settings into your existing ~/.codex/supermemory.json, replacing the placeholder with the key printed by your local server:

{
  "baseUrl": "http://localhost:6767",
  "apiKey": "REPLACE_WITH_LOCAL_SERVER_KEY"
}

An exported SUPERMEMORY_CODEX_API_KEY takes precedence over this file, so unset it if it would select a different key. For server 0.0.9 or newer, also select "apiVersion": "v5"; custom REST URLs otherwise default to legacy compatibility as described above. Keep credential-bearing files private and out of version control.

Keep hosted MCP with a separate key

Codex filters the environment of stdio MCP processes. A fresh plugin install adds only SUPERMEMORY_CODEX_API_KEY to the server's env_vars; exporting SUPERMEMORY_MCP_URL in your shell alone does not forward it. To use a supported alternative MCP service, append SUPERMEMORY_MCP_URL to the existing env_vars list, or set it in [mcp_servers.supermemory.env]. Do not point it at the local REST base unless that service actually implements the expected MCP protocol.

For local REST plus the default hosted MCP, merge this table into ~/.codex/config.toml, replacing the placeholder with a hosted-platform key:

[mcp_servers.supermemory.env]
SUPERMEMORY_CODEX_API_KEY = "REPLACE_WITH_HOSTED_PLATFORM_KEY"

Codex's server-specific env values override inherited/allowlisted values in the proxy process without changing the REST hooks' environment. This supports separate transport credentials without a dedicated MCP-key setting. Preserve the installer's existing command, args, and other fields; edit an existing table rather than adding a duplicate TOML table. Reinstall preserves env and additional env_vars entries. Restart Codex after configuration changes.

Disable MCP for REST-hooks-only memory

Set this field in the existing server table in ~/.codex/config.toml:

[mcp_servers.supermemory]
enabled = false

Keep its other fields intact. Reinstall preserves enabled = false; deleting the server table instead lets the installer recreate it. Automatic REST recall, session profiles, capture, and REST status checks remain available, but explicit MCP memory tools and $supermemory-index writes are unavailable. Turning recallMode off does not disable MCP.

This disables the plugin's hosted MCP memory transport, not every outbound connection. Browser authentication has its own endpoint, session-start checks npm for updates, and Codex or the local backend may use remote services. Supply a local key to avoid automatic browser login; a REST override is not an offline or all-traffic-local guarantee.

Environment variables

~/.codex/supermemory.json (optional)

Drop this file in to override defaults:

Project tags combine the sanitized repository name with a normalized Git-remote hash. Linked worktrees and clones of the same remote therefore share one container; same-named repositories with different remotes do not collide. Without a remote, the Git common directory is used as the fallback identity.

Entity context

Codex sends one shared coding-agent entityContext whenever it saves memories. It covers durable preferences, workflows, architecture, conventions, setup, decisions, and implementation lessons without one save type overwriting another container-level context.

Signal extraction (optional)

When signalExtraction is enabled, only conversation turns containing signal keywords (like "prefer", "decided", "remember", "bug", "fix") are captured. This reduces noise but may miss some context. Disabled by default — all turns are captured.

Commands

npx codex-supermemory install     # set up hooks + MCP + skills
npx codex-supermemory uninstall   # remove hooks + config (keeps your memories)
npx codex-supermemory status      # show current install status

Status

Run $supermemory-status inside Codex to check the saved credential, API reachability, active project container, and account details. Browser authentication is automatic on SessionStart; there is no separate login skill.

Index a codebase

Run $supermemory-index inside Codex to explore the repository and save several focused memories about architecture, conventions, and how to run it. The skill is an agent workflow; the actual writes go through the hosted MCP add_memory tool.

Privacy

Anything wrapped in <private>...</private> is replaced with [REDACTED] before being sent to Supermemory. Use this for secrets, tokens, or anything you'd rather not have stored.

License

MIT