I built one inspired by OpenClaw's approach to file-based memory.

It stores facts in a plain `memory.md` file and chat history in daily logs. On startup, it builds an in-memory semantic index from those files using the embedding model you pass in. The files are the source of truth, embeddings are just a search cache.

`npm install memdir`

https://github.com/artiebits/memdir

It's naive and in beta, but has zero dependencies and is fully local, so no data leaves your machine.

Feedback appreciated.