Hey HN, I'm the author.
I got tired of re-explaining my codebase to Claude and Copilot every session. They forget everything—architecture, patterns, conventions. It's like talking to a goldfish.
So I built In Memoria: an MCP server that gives AI assistants persistent memory. It learns your coding patterns from your actual code (via tree-sitter) and provides that context to any MCP-compatible AI.
Demo (2 min): https://asciinema.org/a/ZyD2bAZs1cURnqoFc3VHXemJx
How it's different from RAG or Cursor rules:
- Not just retrieval: It learns patterns statistically ("90% of your functions return Result<T>")
- Not static config: Learns from actual code structure and usage
- Local-first: Rust core for speed, everything stays on your machine
Since launch: 86 stars, 14 forks, and yesterday we got 2 PRs for PHP support from external contributors I've never met. People are building features I haven't had time to implement, which is wild.
Works with Claude Desktop, Claude Code, Copilot, Cursor, anything that supports MCP. Install via `npx in-memoria server`.
What's the most annoying thing your AI assistant forgets? I'm prioritizing the roadmap based on real pain points.
