PMB

PMB is an open-source, local-first memory system designed for AI coding agents.

Our verdict

Who it's for

Local-first AI memory for coding agents.

Who should look elsewhere

Requires initial setup and configuration for each agent (though simplified with pmb connect commands)

+Fully local and open-source (Apache-2.0), no data leaves your machine, no API keys needed
+Works across multiple AI coding agents (advanced models, Cursor, Zed) without losing context
+Fast hybrid recall (minimal latency) with no LLM call on the read path, keeping latency minimal

−Requires initial setup and configuration for each agent (though simplified with pmb connect commands)
−Focused on AI coding agents; may not be designed for general AI assistant memory (e.g., chat bots outside coding)

✓Automatic memory injection before the model thinks (pre-read) and journaling after actions (post-write)
✓Hybrid recall with BM25, dense vector embeddings, entity graph, and optional reranker fused via Reciprocal-Rank-Fusion
✓Sub-millisecond async writes: returns instantly to agent, embeds and vector storage happen in background thread
✓Memory scoring: tracks whether each lesson is followed by the agent; flags ignored lessons as dead
✓Interactive local dashboard with entity graph map, timeline (git-graph style), and overview
✓No LLM call on the read path; optional summarization and graph extraction can run on local Ollama models
✓MCP integration via stdio, no network, ports, or tokens required
✓Supports multiple projects and agents sharing the same workspace

No public pricing plans recorded.