oleksiijko/pmb

PMB (Personal Memory Brain) gives AI coding assistants like Claude Code, Cursor, and Codex a persistent memory that survives between sessions. Normally these tools forget everything when you close them, so you end up pasting the same project context every day. PMB stores facts, decisions, and preferences on your local disk and makes them available to the AI the next time you open it. It works through MCP, a standard protocol that lets AI tools load external tools and data sources. After installing PMB with pip, you run one command to connect it to your agent: for example, "pmb connect claude" writes the necessary config entry so Claude Code can reach the memory server. You can point multiple tools at the same workspace so Claude Code and Cursor both see the same stored facts. Memory is triggered by explicit phrasing. When you tell the AI to "remember" something, or describe a project decision, PMB stores it. When you ask about past research or a past decision, it retrieves it. General factual questions (like "what is JWT?") bypass the memory layer entirely. You can browse what is stored through a terminal interface or a local web dashboard. The storage uses SQLite and LanceDB on your machine, with no cloud service and no API key required. The cost per retrieval is zero. The project reports 94.5% recall accuracy on a standard benchmark (LoCoMo), with a median response time of 70 milliseconds. It supports retrieval across 50 or more languages, including cross-lingual queries where you ask in one language and the stored fact is in another. The project is Apache 2.0 licensed and installable from PyPI as the "pmb-ai" package. It supports nine different AI tools including Cursor, Codex, Windsurf, Gemini CLI, VS Code Copilot, Zed, and Continue. The full README is longer than what was shown.