EvoGraph-Mem: Editable Graph Memory for Long-Term AI Agents

The research introduces EvoGraph-Mem, a novel framework designed to enhance the long-term memory capabilities of language agents. Unlike existing memory systems that primarily store and retrieve information, EvoGraph-Mem focuses on actively maintaining memory quality by addressing "memory pollution," where outdated or incorrect insights degrade performance over time. This system models insights as nodes in an editable graph, each tracking positive and negative evidence, along with an activation state. EvoGraph-Mem incorporates a utility-aware retrieval mechanism and a graph controller that dynamically updates the memory graph after each task execution. This controller intelligently preserves reliable insights, archives invalid ones, revises outdated information, and integrates newly discovered reusable insights. Extensive experiments demonstrate that this approach consistently outperforms other memory-based agent baselines, highlighting the critical role of active memory maintenance for agents engaged in extended interactions and evolving tasks.