PMB

PMB is a local-first memory system for AI coding agents. It stores decisions, lessons, project facts, and other context in a single SQLite file on your machine. When you start a new session with Claude Code, Cursor, Codex, or Zed, PMB automatically injects relevant memories before the agent even starts reasoning. No cloud, no API keys, no telemetry — everything stays offline.

How does it work? PMB hooks into your AI agent via the Model Context Protocol (MCP). Every message is classified in under a millisecond, and matching lessons, decisions, and project overviews are fetched for the agent before it answers. Writes are async and sub-millisecond — SQLite first, then embedding and vector storage on a background thread. Recall uses a hybrid approach: BM25 text search, dense vector embeddings, and an entity graph, all fused with Reciprocal-Rank-Fusion. The system also scores each lesson by how often the agent actually follows it, flagging dead ones so your memory stays lean and honest.

Who is this for? Anyone who uses AI coding agents regularly and is tired of re-explaining their project every session. PMB is especially useful if you switch between different agents (Claude Code, Cursor, Codex) — they all share the same memory. It's open source (Apache 2.0) and runs entirely on your machine, so you don't have to worry about data leaving your computer. The dashboard includes a Map (force-directed graph of entities) and a Timeline (git-style journal of events), giving you full visibility into what your agent remembers.