Memory that never leaves the box
Why we built agent memory as one local SQLite file, and why "local-first" is a security decision before it's a performance one.
Every cloud memory product asks you to accept the same trade: hand over the most sensitive thing your agent produces — its accumulated knowledge of your users, your cases, your finances — and trust that it stays put. Locamem starts from the opposite premise. Your agent's memory is one SQLite file on your disk. That's the entire system. You can cp it, git it, encrypt it, or delete it. No account, no console, no egress.
This is a security decision before it's a performance one. When the memory never leaves the box, the threat model collapses. There's no provider breach that exposes your data, because the provider never had it. There's no metered API that logs every recall, because there's no API. There's no data-processing agreement to negotiate, because nothing is processed off-machine. For legal, healthcare, and finance teams, "the data physically cannot leave" is an answer that survives an audit. "We promise we encrypt it" is not.
Retrieval runs entirely on-device: SimHash (LSH) for candidate generation, FTS5 for full-text, optional on-device embeddings for semantics. Sub-10ms, zero network calls, zero per-recall cost. We don't meter memory — that's not a pricing tier we're saving for later, it's a thing that cannot exist in this architecture. The engine is MIT-licensed and free forever.
The performance benefits are real and they're a happy side effect. The point is sovereignty. Air-gapped works because there was never a network dependency to remove. The honest comparison: this is what hosted cloud memory stores do, but local-first — same job, except the memory stays on your hardware.
We think this is the only defensible default for agent memory. The burden of proof should be on anyone asking you to ship your knowledge base to their servers, not on the team that lets you keep it. Install is one line: curl -fsSL https://locamem.com/install | bash.