PMB adds local memory for AI coding agents
PMB is a local-first memory tool for . It stores project memories in one file and a local LanceDB vector index, so it does not need a server, cloud service, or API key. Memory lookup combines BM25 search, vector search, and a graph of related entities that appear together.
The goal is to find the memory that is useful for the current task, not just the text that looks most similar. Through MCP, it connects to the agent’s normal work cycle: relevant memories are added before the agent answers, and decisions or new lessons are saved after each turn. The developer does not need to keep typing “remember this.” PMB also builds a project dictionary from stored memories, which helps it find project-specific names and terms later.
With a local Ollama model, optional features such as memory cleanup, duplicate removal, and chatting about stored memories can also stay on the user’s machine.
Key points
- PMB gives a local .
- It uses SQLite and LanceDB, with no required server, cloud service, or API key.
- MCP lets it add memories before an answer and save new decisions after each turn.
- It combines BM25 search and vector search to find useful past context.
- Ollama can keep optional features on the user’s own machine.