pi-claude-mem

A self-maintained memory extension for pi-coding-agent. It is a thin client for the local claude-mem worker HTTP API: it records tool activity to the worker, injects relevant past context each turn, and registers a memory_recall search tool. pi itself does no summarization — the worker at 127.0.0.1:37701 handles summarization, embeddings and vector search.

This is a fork, not an independent reimplementation. It derives from the ArtemisAI [email protected] adapter, which itself derives from the OpenClaw plugin shipped with claude-mem. See NOTICE for the full provenance and the list of modifications.

License

AGPL-3.0-or-later — see LICENSE. The original copyright and notices must be retained. Note this is stricter than the rest of agent-memory-bridge, which is MIT; this agents/pi/ subdirectory is a self-contained AGPL-licensed component.

The claude-mem worker backend is a separate program (Apache-2.0) that this extension only talks to over a local HTTP API; it is neither contained in nor published by this package.

Install as a local-path package

This package is "private": true" and is not published to any npm registry. pi loads the whole directory as a local-path package, so it is never overwritten by pi update.

# 1. make sure the worker is up
curl -s http://127.0.0.1:37701/api/health

# 2. install THIS directory as a local package (run from agents/pi, or pass an absolute path)
pi install "$PWD"

pi install appends the path to the packages array of ~/.pi/agent/settings.json:

{
  "packages": [
    "/absolute/path/to/agent-memory-bridge/agents/pi"
  ]
}

Restart pi. At startup it logs the resolved worker URL, e.g. [pi-claude-mem] worker → http://127.0.0.1:37701 (端口来源: settings.json). Then run the /memory-status slash command — it should report the worker version and degraded: false.

Uninstall

pi remove /absolute/path/to/agent-memory-bridge/agents/pi

then restart pi.

How it works

pi-coding-agent
   └── this extension (extensions/pi-claude-mem.ts)
        session start      local session id / project name
        before agent start -> POST /api/sessions/init        (with prompt)
        context            -> GET  /api/context/inject
        tool result        -> POST /api/sessions/observations (fire-and-forget)
        agent end          -> POST /api/sessions/summarize
        memory_recall tool -> GET  /api/search
        /memory-status     -> GET  /api/health
                              │
                              ▼
                local claude-mem worker (default 127.0.0.1:37701)
                SQLite (FTS5) + Chroma, shared across engines

Every observation is tagged platformSource: "pi-agent".

Configuration

| Variable | Default | Description | |---|---|---| | CLAUDE_MEM_HOST | 127.0.0.1 | Worker host | | CLAUDE_MEM_PORT | 37777 (fallback) | Worker port; the extension also reads CLAUDE_MEM_WORKER_PORT from ~/.claude-mem/settings.json (accepts both number and numeric-string) | | CLAUDE_MEM_DATA_DIR | ~/.claude-mem | Worker data dir (where settings.json is read from) | | PI_MEM_PROJECT | derived from cwd (pi-<dir>) | Project grouping for observations | | PI_MEM_DISABLED | — | Set to 1 to disable the extension |

Resolution order for the port: CLAUDE_MEM_PORT env → CLAUDE_MEM_WORKER_PORT in ~/.claude-mem/settings.json → fallback 37777.

The runtime only imports the two peer modules pi injects (@mariozechner/pi-ai, @mariozechner/pi-coding-agent) plus Node built-ins, so the directory needs no node_modules. The @earendil-works pi distribution exposes compatible extension APIs and loads the package as-is.

Modifying

Edit extensions/pi-claude-mem.ts in place and restart pi — there is no copy to sync (unlike an npm-package patch workflow). Verify with /memory-status.

Files

extensions/pi-claude-mem.ts   extension body (event hooks + memory_recall + /memory-status)
skills/mem-search/SKILL.md    skill that guides the model to use memory_recall
package.json                  private local-package manifest (pi.extensions / pi.skills)
LICENSE                       AGPL-3.0 full text + copyright header
NOTICE                        provenance and modification list

For a from-scratch machine setup (installing pi, the worker and the AI provider) see DEPLOY.md.