Mem0, a personalized AI memory platform, has unveiled its latest research on a long-term memory algorithm that outperforms OpenAI's memory functionality by 26% in accuracy, according to experimental data on the LOCOMO benchmark. The algorithm also significantly reduces P95 inference latency by 91% and cuts token consumption by 90%, addressing AI agents' forgetfulness in prolonged interactions. Mem0's approach involves a two-stage pipeline: an extraction phase that gathers key facts from conversations and historical records, and an update phase that refines these facts using a vector database. This ensures a concise and consistent memory repository. An enhanced version, Mem0ᵍ, uses a graph database to map complex relationships. The system completes memory retrieval and response generation in 0.71 seconds, compared to nearly 10 seconds for traditional methods. The research has been accepted by the European Conference on Artificial Intelligence and is available on GitHub.