New memory method cuts long-term agent storage while improving accuracy
MemoWM: How World Models Change What Agents Need to Remember
Machine LearningArtificial Intelligence
Summary
Long-running AI agents need to remember a lot, which can become very costly in storage. The authors created MemoWM, a method that uses a shared model of the world to keep only the most useful pieces of information while reconstructing other data when needed. This method balances how much space data takes with how important it is for future tasks, leading to better accuracy and much less storage use compared to previous techniques. They also found that stronger world models help reduce stored data even more without hurting task performance. Finally, they highlight a trade-off between the size of the shared model and how much individual experience data is saved.
What this means in practice
- •For autonomous system developers: Reduce memory requirements in long-running autonomous agents by selectively storing experience while maintaining task performance using shared world models.
- •For robotics engineers: Implement efficient memory management for robots operating for extended periods by compressing experience data with predictive reconstruction guided by task relevance.
Authors
Bingfan Zeng, Zhisheng Chen, Chenbo Sang, Zhengwei Xie, Jinpeng Wang, Xiangchen Guan, Rui Qian, Zheng Lu, Jingwei Song
Abstract
Long-term agents face growing storage demands as they accumulate experience. World models capture reusable regularities that can reduce the information stored for each experience. We formulate the problem of memory allocation conditioned on a world model and introduce MemoWM, a framework that uses shared predictions to compress retained information and reconstruct omitted content. Its task-aware allocation rule balances the expected impact of reconstruction errors against storage cost, retaining information with downstream value beyond the predictive prior. Across five long-term agent-memory benchmarks, MemoWM achieves 42.42\% average answer accuracy, exceeding the strongest baseline by 2.62 percentage points, while reducing average experience-specific storage by 53.9\% relative to MIRIX, the most storage-efficient baseline. Further analysis shows that stronger world models reduce per-experience storage at comparable task quality. Accounting for model parameters reveals a trade-off between shared model capacity and recurring storage costs, with the capacity that minimizes total storage increasing as more interactions are retained. Our code is available at https://github.com/Feld-maxiu/MemoWM.