What is PMB?

PMB is an open-source, offline-first memory system that gives AI coding agents like Claude Code, Cursor, Codex, and Zed persistent project memory through the Model Context Protocol (MCP). It stores decisions, lessons, goals, recent work, project facts, and documentation in a single SQLite workspace on your local disk. No cloud, no API keys, and no LLM calls on the read path—everything runs entirely on your machine.

Who it's for

• Solo developers who are tired of re-explaining their project conventions and decisions to AI agents every single session
• Multi-tool teams that switch between Claude Code, Cursor, Codex, and Zed and want their project context to follow them seamlessly
• Privacy-conscious engineers who need persistent agent memory without sending data to any cloud service or requiring API keys

Key features

Auto-recall on every prompt

Every message is classified in sub-millisecond time, and matching lessons, decisions, and project overviews are fetched for the agent before it reasons. The read path takes just 4–16 ms, so memory surfaces without slowing down the conversation.

Sub-millisecond async writes

The MCP tool returns instantly. SQLite writes happen first, while embedding and LanceDB vector index updates run on a background thread, never blocking the agent's turn. This means recording new memories costs essentially zero time.

Hybrid recall with fusion

PMB combines BM25 text search, dense vector embeddings, an entity graph, and optional reranking—all fused with Reciprocal-Rank-Fusion (RRF). One call returns the right memory, ranked, with a 94.5% recall@10 rate.

Honest impact tracking

Every lesson is scored by whether the agent actually follows it. Useful rules get starred automatically; ignored ones are flagged as dead. This lets you prune what doesn't help and keep your memory honest, not bloated.

What stands out

Memory that doesn't wait to be asked—hooks inject the right context before the model thinks, and journal the agent's work after, with no LLM call on the read path.

This is the core difference from other memory tools. PMB doesn't require the agent to remember to call a tool or ask for context. It automatically surfaces relevant memories on every prompt and records new ones asynchronously. The result is persistent context that compounds across sessions without any manual effort.

Worth checking out if…

You want your AI coding agents to actually remember your project decisions, conventions, and past bugs without re-explaining everything each session. PMB is especially useful if you switch between multiple agent tools and want your context to follow you, or if you need memory that stays entirely on your machine with no cloud dependencies.