Mem0 ('mem-zero') is a memory layer for AI assistants and agents. It extracts salient facts from conversations and stores them, then retrieves the relevant ones at inference time so an assistant can remember user preferences, adapt over time, and maintain context across sessions — aimed at use cases like support chatbots, personal assistants, and autonomous systems. It supports multi-level memory across user, session, and agent state.
A 2026 memory algorithm reworked how memories are formed and retrieved: single-pass ADD-only extraction (one LLM call, memories accumulate rather than being overwritten), agent-generated facts stored as first-class entries, entity linking across memories, multi-signal retrieval fusing semantic, BM25 keyword, and entity matching, and time-aware temporal reasoning. Reported benchmark scores (managed-platform figures; open-source results are expected to be directionally similar):
| Benchmark | Old | New | Tokens | Latency p50 |
|---|---|---|---|---|
| LoCoMo | 71.4 | 92.5 | 7.0K | 0.88s |
| LongMemEval | 67.8 | 94.4 | 6.8K | 1.09s |
| BEAM (1M) | — | 64.1 | 6.7K | 1.00s |
| BEAM (10M) | — | 48.6 | 6.9K | 1.05s |
It is available as a library, a self-hosted Docker server, and a managed cloud platform, with Python/Node SDKs, a CLI, and an MCP server. It requires an LLM to operate (default gpt-5-mini) plus an embedding model, and integrates with frameworks like LangGraph and CrewAI. Agents can self-provision an API key that a human owner can later claim.
