Right now, tools like Claude Code, Cursor, and Windsurf start every session with a blank slate. That architectural decision you explained yesterday? Gone. The debugging pattern that took 20 minutes to figure out? You'll explain it again tomorrow.
I kept running into this, so I built Mind Keg MCP, an open-source memory server that gives AI agents persistent knowledge across sessions.
The idea is simple: agents store small, atomic learnings as they work with you. Next session, they pull back what's relevant via semantic search. Knowledge compounds. The agent gets sharper the more you use it.
A few things that made this practical to build and use:
It works with any MCP-compatible agent: Claude Code, Cursor, Windsurf, Codex CLI, Gemini CLI, GitHub Copilot. Semantic search runs locally via ONNX embeddings, so there are no API keys and no costs. Learnings can be scoped per repo, per workspace, or globally. Everything lives in SQLite, zero config, fully offline, and it ships with API key auth and per-repo access control.
Built with TypeScript and Node.js 22.
npm install -g mindkeg-mcp
GitHub: github.com/carloluisito/mindkeg-mcp
Curious what others think. What context do you wish your AI coding tools actually remembered between sessions?
Top comments (0)