PMB

PMB · Personal Memory Brain - local-first persistent memory for AI agents (Claude Code, Cursor, Codex).

Headline numbers

• 94.5% LoCoMo recall@10 (reproducible, full 10-conversation run)
• 70ms p50 warm recall (~3.7s MCP cold start)
• 99.2% top-10 on 900-query multi-language stress test
• 50+ languages via multilingual embedder, with explicit pattern coverage

What's inside

• MCP-native: one command to wire into Claude Code / Cursor / Codex
• Hybrid retrieval: BM25 + vector + PAMVR boosts, no LLM in the hot path
• Auto VOCAB_BRIDGES: domain adaptation mined from your own data
• Pattern query split: compound queries handled via fan-out + RRF fusion
• Atomic fact extraction: mem0-style, regex-based, no LLM
• Keyed-upsert: fact replacement with history (no duplicate current values)
• Durable embed queue: SQLite-backed, survives process restarts
• Multilingual safety: pmb doctor warns if your model doesn't match your data
• Web dashboard at http://127.0.0.1:8765

Install

pip install pmb-ai
pmb connect claude-code

Restart your agent and say "remember - I prefer Postgres". See the README for full benchmarks, screenshots, and architecture.

Compared to alternatives

| | PMB | mem0 | Letta | Zep |
|---|---|---|---|---|
| LoCoMo recall@10 | 94.5% | ~67-70% | ~76-80% | ~80% |
| p50 latency | 70ms | 1-3s | 1-3s | 1-3s |
| Runs offline | yes | no (cloud) | partial | partial |
| API key required | no | yes | yes | yes |

Honest limitations

• Top-1 universal ceiling without LLM rerank: ~58% on LoCoMo, ~73% on the multi-language stress. Optional recall.llm_rerank (Ollama) lifts top-1 by 10-15pp at +200ms latency cost.
• Some adversarial cases (mixed-language queries with ambiguous proper nouns) remain hard without LLM-based coreference.
• Tested primarily on Windows + Linux. macOS works but has fewer CI runs.

License

Apache 2.0. No telemetry, no API keys, no cloud.

Full Changelog: https://github.com/oleksiijko/pmb/commits/v0.1.0