Fleet Memory
Self-hosted shared memory for a team of AI agents. Rooms, L0-L3 depth, no LLM on the read path.
RCLL is a shared memory for a fleet of agents, where every fact carries a room, a hall, and a layer, and recalls can request exactly the needed slice. It can be fully self-hosted: storage and retrieval run locally, with no account, no API key, and no outbound call required to read. The read path never calls a model; it uses only a 33M-parameter embedder and an 80 MB reranker, both local by default.
Every fact is assigned to a room, and a recall can filter to specific rooms, returning zero results if the room does not contain the answer. The default retrieval configuration is not optimal for multi-hop questions; fusion plus reranking scores lower than plain dense plus reranking.
Summary drafted from the project's own website. Every sentence is backed by text on that page and was reviewed before publishing.