https://arxiv.org/abs/2506.15841

The research introduces MEM1, a novel reinforcement learning framework designed to enhance language agents' efficiency and performance in complex, multi-turn interactions. Unlike traditional models that accumulate information, MEM1 uses a constant-memory approach by integrating prior knowledge with new observations into a compact internal state, strategically discarding irrelevant data. This method significantly reduces computational costs and memory usage while improving reasoning, particularly in long-horizon tasks such as question answering and web navigation. The authors also propose a scalable task augmentation strategy to create challenging multi-objective environments, demonstrating MEM1's ability to generalize beyond its training horizon and exhibit emergent, sophisticated behaviors.