pi-mnemo

Token-frugal long-term memory for pi. Adds ~600 input tokens per request (two tools + one guideline) instead of ~10k for heavier memory extensions.

Install

pi install npm:pi-mnemo

On the first interactive start pi downloads the search model (~55 MB) in the background. Until it arrives, memory still works with text search only.

How it works

• Storage: ~/.pi/agent/mnemo/mnemo.db (built-in node:sqlite, WAL, shared by all pi processes).
• Memories are English notes (fact, preference, decision, pitfall, procedure), global or per project (git root name). Each has English tags and a hidden Russian search field (ru) so Russian prompts still match English memories.
• Search: FTS5 over Russian/English Snowball stems + static multilingual embeddings (static-similarity-mrl-multilingual-v1, 512 dims, int8) read row-by-row from disk, fused with reciprocal rank fusion. ~2 ms per query, no resident model.
• Auto-recall: on every non-trivial user prompt, up to 3 memories that pass a calibrated gate are injected as a <memory> message. Memories already shown in a session are not repeated. In the TUI it shows as a 🧠 memory block with one line per note (id, kind, snippet); click it or press ctrl+o to see full text, date and project.
• Error hints: when a tool fails, the error is matched against stored memories (stricter gate, at most once per 10 tool calls) and a hint is appended to the tool result. The TUI also shows it as a separate 🧠 similar issue seen before block (display only, not sent to the model).
• Background writer: checkpoints after each agent run, every 25 tool calls or ~60k chars during long runs (at most every 5 min), before compaction and at shutdown. A job reads the new slice of the session file, shows claude-haiku-5-5 (fallback gpt-6-luna) the slice plus related memories, and applies add / update / supersede operations and a rolling session digest. Jobs live in SQLite with leases, so they survive restarts and run in any interactive pi process.
• Tools: recall({queries[], scope?, kinds?, sessions?, limit?}) and remember({text, kind, scope?, replaces?}). Both return structuredContent, so codemode scripts get plain objects; usage notes live in the mnemo namespace instructions (describeNamespace("mnemo")).
• Secrets are redacted before anything is sent to a model or stored.

Commands

/mnemo off | on | stats | search <q> | show <id> | forget <id> | import-hermes [enrichment.json] | backfill | enrich

/mnemo off pauses memory for the current session: no auto-recall, no error hints, recall/remember are removed from the active tools, and everything recorded while paused is skipped by checkpoints, digests and backfill. The state is saved in the session and restored on resume; /mnemo on resumes.

CLI (works from the installed package too): npx pi-mnemo fetch-model | import-hermes | backfill | search <q> | stats. Inside pi's package directory: node ~/.pi/agent/npm/node_modules/pi-mnemo/bin/pi-mnemo.mjs <command>.

Environment

Subagent children (PI_SUBAGENT_CHILD) can recall but never run the background writer.

Development

npm ci
npm run fetch-model
npm test

eval/bench.ts replays the retrieval benchmark. Its data (real memories and queries) is local only and not committed.

Releasing

CI (.github/workflows/publish.yml) publishes to npm when a v* tag is pushed. Publishing uses npm trusted publishing, so no token is stored in GitHub.

npm version patch        # bumps package.json, commits, creates tag vX.Y.Z
git push --follow-tags   # the workflow tests, checks the tag matches, publishes