MIA, a next-generation memory agent framework, aims to end agents' "amnesiac" workflows
新一代记忆智能体框架MIA:让智能体告别「失忆式工作」,在持续进化中变强
A Shanghai Institute for Advanced Learning and ECNU team released MIA, a memory agent framework, and said it achieved the best results on 7 datasets. MIA uses a Manager-Planner-Executor design, dual parametric and non-parametric memory, alternating RL, and test-time continual learning; the post does not disclose exact benchmark scores. The key point is memory as capability internalization, not just retrieval, for open-world agents.
Why it matters: A Shanghai research team's MIA memory framework splits agent memory into parametric and non-parametric types and updates them with alternating reinforcement learning, offering a transferable way to keep evolving on open tasks.