AI tools can do the work. 3ngram helps them keep the thread.

AI work is increasingly conversational, but conversations are temporary. A decision made in Claude, a commitment captured in ChatGPT, and a blocker explained to Codex should not disappear when the tab closes. 3ngram gives those tools one typed memory layer over MCP.

I built 3ngram because every AI tool I use gave me a different place to think, while I remained responsible for carrying the context between them. The model was never the durable object. The decisions, commitments, facts, and working preferences were.

Capture is explicit. You or your agent chooses what to remember. Every memory has a type, scope, project, and lifecycle. Corrections append a new version and preserve the old one, so the system never silently rewrites what you believed.

The memory engine, MCP server, REST API, SDK, CLI, and self-host backend are Apache-2.0. The managed dashboard, billing, and cloud operations are proprietary. That boundary lets you inspect and self-host the memory layer without pretending the hosted product is the same thing as the public repository.

Right now, 3ngram is one person in Copenhagen. That's on purpose. Founded 03.03.2026, building slowly and talking to every early user myself. Hiring will come when the product earns it, not before.