Mem0#
Open-source memory management platform for LLM agents. Provides selective memory extraction, graph-enhanced retrieval (Mem0g), and multi-scope memory (user, session, agent, run, org). The most benchmarked memory system in the wiki.
Key Features#
- Two-phase pipeline: extraction (discrete facts, not summaries) → update (ADD/UPDATE/DELETE/NOOP)
- Graph-enhanced retrieval (Mem0g): entities as nodes, relationships as edges. 68.4% accuracy vs 66.9% vector-only.
- Five scoping dimensions: user_id, session_id, agent_id, run_id, org_id
- Forgetting: active decay based on Bjork’s Theory of Disuse
- 93% token reduction vs full-context approach (1,800 vs 26,000 tokens)
Benchmark (LOCOMO)#
Best-in-class among memory systems: Mem0g 68.4%, A-Mem 68.6%, full-context 72.9% (but at 14× token cost and 7× latency).
Integration#
LangChain/LangGraph (retriever + storage backend), Mastra (remember + memorize tools), MCP-compatible clients.
Position in Landscape#
Mem0 represents the Graph + Vector pattern — the recommended approach for most autonomous agents per agent-memory-systems-2026. Contrasts with pai’s three-tier memory (hot/warm/cold) and the llm-wiki-pattern’s file-based approach.