Source: Mem0 documentation, open source overview and Platform vs OSS

Mem0's docs on deployment modes: OSS defaults (gpt-5-mini, text-embedding-3-small, local Qdrant, SQLite history), the self-hosted server, Platform-only features, and the removal of external graph stores.

Contents

Summary

Mem0 ships as an open-source library (pip install mem0ai, npm install mem0ai), a self-hosted server, and a managed Platform. The OSS library is the extraction and retrieval core; many lifecycle features are Platform-only.

Key claims

  • OSS library defaults: LLM OpenAI gpt-5-mini, embedder text-embedding-3-small, vector store local Qdrant at /tmp/qdrant, history in SQLite at ~/.mem0/history.db, reranker off until configured.
  • Self-hosted server (make bootstrap in server/, Docker Compose): Postgres with pgvector, dashboard, per-user API keys, request audit log; bundled LLM providers OpenAI, Anthropic, Gemini.
  • Both offer async clients (AsyncMemory OSS, AsyncMemoryClient Platform), reranking and multimodal input.
  • Graph memory: the Platform has a native always-on entity graph that feeds ranking. Links come from entity co-occurrence, not typed relations; the old relations response field is now always empty. The earlier external graph-store integration (Neo4j, Memgraph, Kuzu, Apache AGE, Neptune via enable_graph and graph_store) was replaced by the native graph on the Platform (enable_graph=True now only adds an empty relations key) and removed from OSS in the v3 pipeline; OSS still extracts entities for ranking but has no queryable graph.
  • OSS has no batch operations (loop over single update / delete).
  • Platform-only: webhooks, schema-driven export, batch update/delete (up to 1000 per call), feedback, summaries, Dream consolidation, memory decay, temporal reasoning, custom categories, dashboard.
  • Platform memories can be scoped by user_id, agent_id, app_id and run_id; OSS has no app_id.

Feeds

Mem0 · Letta

Suggest changes →