Today I published both versions on npm — the MCP version and the Engine version. I still have too little karma for Show HN (T^T)
This is the same engine I built for the hackathon. I kept using it after submission, and during the judging period the conversation grew to ~480 turns in a single session.
With my wrapper configured to keep an active window of roughly 6–8 recent turns, model input usually stays around ~2,500–15,000 tokens per turn. The lower end is normal conversation; it grows when images, attachments, or larger context blobs are still inside that active window.
*The active window is configured by the wrapper, not the memory engine itself.
https://www.npmjs.com/package/@nature-labs/living-memory-engine
⸻
The MCP version serves a different purpose, but uses the same engine underneath.
I think of it like spawning a small moon that orbits outside your agents. Any agent in the same embedding space (same dimensions, e.g. 768 or 1024) can write to and recall from the same memory world.
That means Claude Code, Claude Desktop/Cowork, Codex, ChatGPT, or any agent that supports MCP can remember and recall the same things together.
You don’t have to tell the same story every day to every agent just because you use them for different sides of your work.
https://www.npmjs.com/package/@nature-labs/lme-mcp
Both are ready to use today. Try them however you like — and if you find something broken or interesting, please open an issue on GitHub --/--
Log in or sign up for Devpost to join the conversation.