A universal memory layer for AI Agents providing multi-level personalized long-term memory extraction, storage, and fused retrieval.
Mem0 ("mem-zero"), incubated by Y Combinator S24, is a universal memory layer for AI Agents. Key highlights include: three-level memory management (User / Session / Agent); a single-pass ADD-only extraction algorithm that completes memory extraction in one LLM call without overwriting existing memories; built-in Entity Linking to establish relationships between memories; and a retrieval layer fusing semantic search, BM25 keyword matching, and entity matching in a single call. The project provides Python SDK, Node.js SDK, and CLI for integration, with three deployment modes: open-source library embedding, Docker Compose self-hosting, and cloud platform. It achieves strong results on memory benchmarks including LoCoMo (91.6), LongMemEval (93.4), and BEAM 1M (64.1). Ecosystem integrations cover LangGraph, CrewAI, Vercel AI SDK, along with coding assistant plugins for Claude Code, Cursor, and a Chrome browser extension.
Quick Start:
pip install mem0ai
pip install mem0ai[nlp] # BM25 + entity extraction
Core Usage:
from mem0 import Memory
m = Memory()
m.add("User prefers dark mode", user_id="alice")
m.search("alice's preferences", user_id="alice")
Unconfirmed Details: The "New Memory Algorithm (April 2026)" date is in the future and unverified; the default LLM is listed as gpt-5-mini, which has not been officially confirmed by OpenAI and may be a placeholder; cloud platform pricing and underlying vector store options are not specified in public materials.