How does retrieval work inside the daemon?

The daemon's local retrieval runs entirely on CPU: a BM25-class lexical lane fused with graph signals over your memory store. That combination covers most agent-memory workloads with zero external dependencies.

Dense retrieval is bring-your-own-embedder: point the daemon at Ollama, vLLM, TEI, llama.cpp, or LiteLLM and the dense lane fuses in alongside lexical and graph. The local dense lane is never metered.

Richer lanes (entity, event, navigational) run on hosted CoreCrux for teams that want them; they are additive, not required.

Every retrieval call, local or hosted, takes a hard token_budget and the daemon trims results to fit. On the ScoreCrux benchmark, budgeted retrieval returns far fewer tokens than naive top-K at equal or better answer quality.