// the find
mem0ai/mem0
Universal memory layer for AI Agents
Mem0 is a memory layer for AI agents that extracts facts from conversations, stores them as embeddings, and retrieves them for future context. It solves the real problem of LLMs forgetting everything between sessions, with a managed cloud option for zero-ops use and self-hosted for teams that can't send data out. The April 2026 v3 algorithm jump is substantial — +27 points on LongMemEval is not a rounding error.
The new ADD-only extraction model is architecturally cleaner than the previous UPDATE/DELETE approach — no tombstone logic, simpler consistency guarantees, and the benchmark numbers back it up. Multi-signal retrieval (semantic + BM25 + entity linking) is the right call; pure vector search misses too many keyword-sensitive queries. The three deployment tiers (pip library, self-hosted Docker, managed cloud) cover the actual range of teams that need this without forcing everyone onto the paid platform. The evaluation framework is open-sourced separately so you can reproduce the benchmark claims rather than just trust the marketing table.
The library defaults to OpenAI for both LLM and embeddings, which means your memory layer bills a second provider even if your main app uses something else — the config to swap this out exists but isn't front-and-center. The ADD-only model that makes v3 cleaner also means memory sets grow forever; there's no documented compaction or pruning strategy, which will bite anyone running a long-lived production agent. Self-hosted auth was off by default until recently and the upgrade notes bury the 'set ADMIN_API_KEY or everything is open' warning — easy to misconfigure. The Python-first SDK means TypeScript users get a thinner wrapper and the npm package lags on features.