Knowell: a hybrid code intelligence and memory engine

A feature can cross an API, an event consumer and a shared package. Its business name may never appear in the code, and the decisions behind it can disappear between agent sessions.

I built Knowell, an open-source hybrid code intelligence and memory engine in Rust, to help coding agents work across multi-repository systems. Agents access it through MCP.

Knowell combines exact symbol and path matches, BM25 text search, optional semantic vectors and evidence-carrying code relationships. It retrieves relevant code from indexed projects and packs it within an estimated token budget, with source versions, freshness and coverage gaps attached. The model’s context limit still applies; the engine helps choose what belongs in it.

Versioned decisions and task checkpoints give a fresh session a place to resume. Continuity comes from records saved to Knowell, rather than an agent’s full conversation history. Agent findings remain proposals until a human accepts them, and records become stale when their evidence changes. A shared hub supports access to saved work from another machine.

The goal is simple: spend less time explaining the same system, and give agents a clearer basis for their work.

Knowell is pre-1.0. The repository includes the implementation and its current limits; the architecture notes describe the full 1.0 target. The source is available under MIT or Apache-2.0. If you build something with it, I’d like to hear about it.