[13pixels]

Very cool to see a local-first approach to agent memory. The shift away from hosted vector DBs for single-agent use cases makes total sense.

Are you using `sqlite-vec` under the hood for the embeddings or a custom extension? Also, curious how you handle the 'hybrid' search part—is it just a linear combination of FTS5 bm25 and vector cosine similarity, or do you have a re-ranking step?