pi-mem-cc

Automatic observer-based memory for pi coding agent.

Inspired by claude-mem for Claude Code, but built natively for pi's extension API. Silently watches every tool call, compresses it into a structured observation, stores it in SQLite, and re-injects relevant past context at session start.

What's different from pi-memory (jayzeng)?

| | pi-memory | pi-mem-cc | |---|---|---| | Memory capture | Manual — agent/user calls memory_write | Automatic — every tool call observed | | Storage | Plain markdown files | SQLite + FTS5 | | Search | Optional qmd (external dep) | Built-in hybrid (keyword + recency) | | Injection | Agent must call memory_read | Auto-injected at session start | | Use case | Personal knowledge book | Cross-session continuity |

Use both. They cover different needs.

Install

pi install npm:pi-mem-cc

Then /reload.

Tools registered

| Tool | Purpose | Token cost | |---|---|---| | mem_search | Get compact index of matching observations | ~50-100 tok/result | | mem_timeline | Get chronological context around an observation id | variable | | mem_get | Fetch full observation details by IDs | ~500-1000 tok/each |

3-layer progressive disclosure — search first, timeline for context, get only for filtered IDs.

Storage

~/.pi/agent/memory/pi-mem-cc.db   # SQLite (observations, summaries, FTS5)

How it works

1. session_start → register session in DB
2. before_agent_start → query top-K relevant past observations (recency × project match), append markdown to system prompt
3. tool_result → fire-and-forget compress into <observation> via observer LLM, write to DB
4. agent_end → fire-and-forget compress turn into <summary> via observer LLM, write to DB

The observer LLM uses the same model as the active session — no extra API key needed.

Configuration

No configuration. Uses your active session model.

License

Apache-2.0