Why agent memory should stay local, what our benchmarks actually mean, and how Locamem is built — written plainly, with the numbers attached.
Why we built agent memory as one local SQLite file, and why "local-first" is a security decision before it's a performance one.
We hit 99.4% recall@10 on LongMemEval-S and ~58% end-to-end QA. The gap is the whole story, and we won't paper over it.
Every recall returns a breakdown — lexical, SimHash, recency, validity — instead of one opaque relevance number.
When a fact changes, the old one isn't erased — it's marked superseded, with the contradiction logged and queryable.
How Locamem plugs into any MCP client with no daemon to babysit and no service to provision.