PMB | Local-first memory for AI

PMB classifies each message in under a millisecond and fetches the matching lessons, decisions and a project overview before the agent reasons. Nothing depends on the model remembering to call a tool. The read path carries no LLM call, so recall stays cheap and fast. You don't manage it. It just happens.

Search fuses BM25 keyword matching, dense vectors and an entity graph, with an optional rerank step, combined through Reciprocal-Rank-Fusion. According to the vendor, this lands at 94.5% recall@10. In practice you get one ranked list instead of guessing which retrieval mode to trust. One call, one answer.

The MCP tool returns instantly. PMB writes to SQLite first, then runs the embedding and LanceDB insert on a background thread. That means recording a decision never stalls the turn you're in the middle of.

Every rule is tracked by whether the agent actually follows it, and useful lessons get starred while ignored ones get flagged dead so you can go back and remove them. The follow-rate readout is the honest part of the system. It tells you which memories aren't earning their keep.

The local dashboard renders your memory as a graph where facts, files and entities are nodes sized by importance, plus a git-style timeline of decisions and events newest-first. Hovering a node dims the rest and lights up neighbors. It's a real view served from your machine, not a mockup. Want to see how your project's knowledge connects? Hover a node and the graph answers.

Claude Code, Cursor, Codex and Zed all read from the same store. Switching agents costs you nothing because your context lives with the project. That's the feature that makes the layer useful rather than just tidy.