Concept
Key concepts
MemLab is a persistent memory layer for LLM and agent applications. You write raw interactions as Episodic Memory; MemLab extracts them into Semantic Memory and a bi-temporal knowledge graph, then serves them back through hybrid retrieval. These four concepts cover the full lifecycle.
The write path
Your app sends an episode (a chat turn, document, tool trace, or event). MemLab stores the raw content, then asynchronously extracts facts, entities, and relations from it. Extraction is eventually consistent — a write becomes retrievable within seconds, not on the same request.
The read path
Retrieval fans out across vector, graph, and lexical backends in parallel and fuses the results with Reciprocal Rank Fusion (RRF). You get the strongest signal across all three, scoped to your organization and project.
Tenancy
Every memory is scoped to an organization and a project. Vector collections, graph data, and relational rows are all partitioned by that pair, so retrieval never crosses tenant boundaries.