GitHub
Local-first memory for your AI coding agent
Start new thread

GitHub - Local-first memory for your AI coding agent

by•
Your AI coding agent forgets your project every session. Most memory tools store your context on their servers. PMB keeps it on yours. It captures decisions, lessons and facts as you work and feeds the relevant ones back automatically via MCP hooks, so your agent shows up already knowing the project. Different: 100% local (one SQLite file, no cloud, no API keys), automatic (no "remember" step), cross-agent (Claude Code, Cursor, Codex), hybrid offline retrieval.

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt 👋 I build with AI coding agents every day, and the same thing kept getting me: every new session the agent forgets everything. I'd re-explain the same decisions, conventions and dead-ends I'd already gone over that week. The hosted "memory" tools solved it by putting my private project context on someone else's servers, which I didn't want. So I built PMB: a memory layer that's local-first and actually mine. It captures decisions, facts and lessons as I work and feeds the relevant ones back to the agent automatically over MCP, so it shows up already knowing the project. Everything lives in one SQLite file on my machine, no cloud, no API keys. The hard part turned out to be not storing memory, but surfacing the right memory at the right moment without slowing the agent down. That's where most of the work went: hybrid retrieval, automatic capture via hooks, and a per-project lexicon that sharpens over time. Open source (Apache-2.0), just hit 1.0, works across Claude Code, Cursor and Codex. Would love your honest feedback, especially if you've tried other memory setups. What's worked for you?