Hey everyone,
Like a lot of you, I’ve been using Cursor heavily, but I got frustrated with it “forgetting” context across different days or when I switch between related projects. Constantly re-pasting docs or relying on simple vector search wasn’t cutting it—it felt like a filing cabinet, not a brain.
So, I built an MCP server called Mnemosyne to fix it.
It’s a local, associative memory backend built in C# and SQLite. Instead of standard RAG, it uses spreading activation and Hebbian decay to simulate how human memory actually works:
- It uses SQLite FTS5 for initial retrieval.
- It performs Breadth-First Search (BFS) across a localized graph to spread activation energy to related concepts.
- Memories you use frequently form stronger connections.
- Unused trivia naturally fades away over time.
It runs entirely locally on your machine (zero cloud dependencies, no vector DB subscriptions).
I just released the standalone binaries (Windows & Linux) in Early Access for a one-time $29 to fund further development (I’m planning to add direct Git repo ingestion next).
You can check it out here: Mnemosyne MCP by Tater Labs — Persistent Memory for Claude & Cursor
Would love to hear what the community thinks of the Hebbian approach vs. standard vector databases, or if you have any feature requests for Cursor-specific workflows!