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

An assistant that forgets you between sessions recommends chicken to the vegetarian who told it so last week. The obvious fix, sending the whole conversation history with every request, gets slower and more expensive with every message, and models use long histories unevenly anyway.

Mem0, from the company of the same name, is a memory layer that sits beside the conversation. After each exchange, an LLM reads the new messages, a running summary of the conversation and a few recent turns, and extracts salient facts. For each candidate fact, the system retrieves the most similar stored memories and asks the LLM to choose one of four operations: ADD a new memory, UPDATE an existing one with new detail, DELETE one the new fact contradicts, or NOOP when nothing changes. At question time, only the relevant memories go into the prompt. A graph variant, Mem0g, additionally extracts entities and relationships into a graph database to help with temporal and relational questions.

On LOCOMO, a benchmark of very long multi-session conversations (about 26,000 tokens each), Mem0 scored higher than other memory systems, including OpenAI's ChatGPT memory, LangMem, Zep and A-Mem, on most question types. The comparison that matters most is buried in the second table: sending the full conversation scored highest of all (72.9 against Mem0's 66.9 and Mem0g's 68.4). What Mem0 bought was speed and cost: 91% lower p95 latency and over 90% fewer tokens per answer.

Customers ask for "an assistant that remembers" constantly. This paper is a useful design reference and a lesson in reading vendor benchmarks.

Memory is a cost decision before it is a quality decision. On this benchmark, full context was more accurate. Memory was what made it fast and cheap: 1.4 seconds p95 instead of 17, and about 1,800 tokens per answer instead of 26,000. For short histories, sending everything may be the better product. Memory earns its place when histories grow past what you can afford to send, or past what fits.

Maintain memories, do not just append them. An append-only store ends up holding "lives in Lisbon" and "lives in Porto" side by side, and the model has to guess which is current. The update and delete operations are the core idea. They are also where errors happen: a wrong DELETE silently removes something true.

Read who wrote the benchmark. This paper is by the company that sells Mem0. The results may well hold, but the comparison choices, such as running ChatGPT memory through its consumer interface with gpt-4o-mini, and the headline emphasis on "26% better than OpenAI" rather than "six points below full context", are the authors' choices. Rerun the comparison on the customer's own conversations.