About Locamem — Memory that never leaves the box

Memory that never leaves the box.

Locamem is local-first memory for AI agents. Your agent's whole memory is one SQLite file on your disk — no servers, no API keys, no context shipped to anyone. You own the file: read it, copy it, diff it, delete it. Retrieval runs on-device in under 10ms with zero network calls and zero per-recall cost.

Remember without giving up control.

Agents should remember without giving up control of what they remember. So we built memory as a single local file. SimHash (LSH) for candidate matching, FTS5 for full-text, optional on-device embeddings for meaning — all running against one SQLite database that lives on your machine. Every result returns a per-facet score breakdown, so a recall is something you can audit, not a black box you have to trust. Local-first isn't a deployment option for us. It's the design.

The problem

Cloud memory means shipping your context to someone else.

Agents are stateless. Every session starts from zero, so the obvious fix is to bolt on a memory layer. Today that almost always means a cloud service: your agent's context — the prompts, the user data, the accumulated history of what it knows — gets shipped to someone else's database, indexed on their hardware, and billed back to you per recall. You inherit their retention caps, their pricing tiers, their outages, and their answer to the only question that matters: who can read your agent's memory? For legal, healthcare, and finance teams, "it's in a vendor's cloud" isn't an answer they can ship.

Our bet

Transparency and temporal truth beat opaque vector stores.

The best memory is the one that never leaves the box. We're betting that two things win over opaque, metered vector stores: transparency and temporal truth. Transparency means every recall comes with its own scorecard — you see exactly why a result ranked where it did, per facet. Temporal truth means memory tracks validity over time and detects contradictions instead of silently overwriting; Locamem never deletes, so you can ask not just what the agent knows but what it knew, and when. A memory you can explain and a history you can replay beat a similarity score from a system you can't inspect.

Five commitments behind the engine

Local by defaultOne SQLite file on your disk. Zero network calls on recall. Air-gapped capable by design, not as a premium add-on.

$0 per recallWe never meter memory and there is no per-recall cost. The local engine is free forever. Storing and recalling more never costs more.

Auditable, not a black boxEvery result returns a per-facet score breakdown across SimHash, full-text, and embeddings. You can see why anything matched.

Never-delete temporal truthMemory carries temporal validity and surfaces contradictions instead of overwriting them. The history stays intact and replayable.

MIT open sourceThe engine is open source under MIT. Read the code, fork it, run it air-gapped. No lock-in, because the file and the engine are both yours.

Built by Wilcoe.

Locamem is built by Wilcoe. We make developer infrastructure for teams that need their agents to remember without handing their context to a third party — which is why we monetize support, optional self-hostable encrypted sync, and enterprise on-prem help, never the memory itself.

The engine is open source on GitHub at github.com/TeamWilcoe/locamem. Questions, security review, or enterprise deployment: team@wilcoe.com.