# Mem0

Mem0 is a memory layer for AI agents that stores user, session and agent-level facts so an assistant keeps context and personalization across conversations, available as a library, a self-hosted server, or a managed cloud platform.

A newer single-pass, add-only extraction algorithm is reported to raise scores on the LoCoMo and LongMemEval memory benchmarks and to hold up at a million and ten million tokens on a separate large-scale benchmark, by treating agent-confirmed facts as first-class, linking entities across memories, fusing semantic, keyword and entity retrieval, and reasoning about which dated memory answers a time-sensitive question. The evaluation framework behind these numbers is open source so results can be reproduced.

Use cases described include customer-support recall of past tickets, consistent multi-session AI assistants, and adaptive personalization for other products. An agent can register its own working API key directly from the command line, with a human able to later claim that account by email without losing existing memories.

Python, Apache-2.0.

Website: https://mem0.ai