Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

Summary

Mem0 incrementally extracts facts from message pairs (summary + last m=10 turns) then ADD/UPDATE/DELETE/NOOP against top-s=10 similar memories via GPT-4o-mini tool calls. Mem0^g stores entity–relation triplets in Neo4j. On LOCOMO (10 convos, ~600 turns / 26k tokens, ~200 Qs each), Mem0 J 66.88 overall vs OpenAI memory 52.90 / Zep 65.99 / full-context 72.90; single-hop J 67.13 and multi-hop 51.15 lead. Graph variant best on temporal (J 58.13) and competitive open-domain (75.71 vs Zep 76.60). p95 total latency 1.44s (91% below full-context 17.1s) at 1764 retrieved tokens vs 26k. Code https://github.com/mem0ai/mem0; research https://mem0.ai/research.

Keywords

mem0 · agent-memory · locomo · graph-memory · rag · neo4j · latency

Topics

agent memory, long-term memory, graph memory, conversational agents

Research notes

• Primary: arxiv abs (cs.CL; also cs.AI). License not stated on abs/HTML at check. Authors at Mem0 (research@mem0.ai). Code https://github.com/mem0ai/mem0 (63,230 stars at check). HF paper page 71 upvotes; githubRepo linked; 3 unofficial linked datasets (survey PDFs / 12-row locomo results) not copied into hf_* fields and not substantial, so no datasets_local row. Evaluates public LOCOMO rather than releasing a new corpus. Discord posted abs plus https://mem0.ai/research. License field left blank per catalog convention.