PMB

PMB's core design decision is that memory should be a file, not a service. Every event lands in SQLite; vectors sit in LanceDB next to it; copying your whole memory is a cp command. That is a meaningful contrast with memory products that store your project context in someone else's database and bill per recall. The retrieval design is what gives PMB a right to exist beyond that. It fuses BM25, dense vectors and an entity graph with Reciprocal-Rank-Fusion, then offers an optional rerank pass. Reads land in 4-16 ms and the vendor reports roughly 35 ms end-to-end recall in a real workspace. When you are mid-edit and waiting on context, that difference is the whole experience. Vendor-published figures put it at 94.5% recall@10 with a 0.774 MRR. The hook architecture is the second real design choice. PMB does not rely on the agent deciding to call a memory tool; the server classifies every prompt in sub-millisecond time and injects the matching lessons, decisions and project overview before the model reasons. Writes go to SQLite first while embedding and the LanceDB insert run on a background thread, so the MCP call returns in under a millisecond and the turn is never blocked. The MCP server itself exposes 29 tools over stdio as a child process of your agent — no network, no port, no token. Strengths: genuinely local (offline, no API keys, no telemetry), Apache-2.0 and pure Python across macOS, Linux and Windows, a Map entity graph and git-style Timeline that make memory legible, lesson honesty scoring with dead-rule pruning, automatic decay, archiving and dedupe, multi-project lanes, and an optional local Ollama path for summarization and graph extraction that keeps the whole loop offline. Four agents — Claude Code, Cursor, Codex and Zed — can read the same workspace, so switching editors costs you nothing. Weaknesses are structural rather than fixable. It is local-first by design, and the vendor documents no cloud sync or multi-user collaboration across machines. Setup assumes you are comfortable with a CLI and MCP configuration. It connects to MCP-compatible agents, and the vendor names Claude Code, Cursor, Codex and Zed. Agents on remote boxes or in CI cannot easily host a local stdio child process. And the benchmark numbers, including 94.5% recall@10, are measured on PMB's own engine and published by the vendor. Where it fits: a developer running a coding agent every day who is tired of pasting the same conventions into a fresh chat, and who wants that context to follow them between Claude Code, Cursor, Codex and Zed. Where it doesn't: a team that needs one shared memory across several machines, a non-technical user who wants a turnkey product, or anyone whose agents live on remote infrastructure.