Source: Mem0 documentation, memory concepts

Mem0's docs on how add extracts facts into SQL, vector and entity stores, and how search ranks memories by semantic, keyword, entity and temporal signals.

Contents

Summary

Describes Mem0’s current storage and API model. add takes conversation messages and stores extracted facts, not the transcript. search blends several ranking signals and should always be scoped to a user, agent or run.

Key claims

  • Three storage layers: a SQL database (facts and metadata, the source of truth), a vector database (embeddings) and an entity store that boosts memories sharing entities with the query. On the managed Platform the entity store also backs Graph Memory; the docs state the open-source library has no graph memory.
  • The add pipeline: look up related existing memories, extract facts with an LLM, deduplicate and embed, extract entities.
  • Extraction is additive in the current docs; corrections need explicit update or delete calls.
  • infer=False stores content verbatim; inference (fact extraction) is the default.
  • Search ranks with four signals: semantic similarity, keyword match, entity match and temporal metadata.
  • user_id, agent_id and run_id filters separate memories between users, agents and sessions; custom metadata (tenant, category) can also filter.

Feeds

Agent memory

Suggest changes →