MemSteer optimizes agent memory using questions from user histories
MemSteer: Turning User Histories into Supervision for Agent Memory Optimization
Artificial Intelligence
Summary
Building computer agents that remember and use past information well is hard because it's expensive to label data to teach them. The authors created MemSteer, a method that automatically turns user history into questions and answers to guide how these agents organize and retrieve memory. They show that using this automatically created supervision helps the agents perform better on tasks, without changing their main learning model. This approach keeps improving memory design as new user history comes in.
What this means in practice
- •For ai system engineers: Enhance memory stores in AI agents by generating supervision from interaction data to improve task accuracy without retraining models.
- •For customer support platform developers: Improve virtual assistant responses by optimizing memory retrieval using automatically generated question-answer pairs from user logs.
Authors
Jie Huang, Cuiling Lan, Qize Yu, Xiulian Peng, Yan Lu
Abstract
An effective agent memory system should adapt to a user's history and task requirements. Human-annotated supervision can guide this adaptation by providing feedback for evaluating and improving memory designs. Yet suitable labels are rarely available for new user histories, and manually constructing QA pairs and supporting evidence is costly. We introduce MemSteer, which turns unlabeled user histories into question-answer (QA) supervision with supporting evidence for memory-design optimization. MemSteer constructs an evidence-grounded answer plan that specifies how facts across related records jointly support a reference answer. The plan guides question generation, while source-grounded auditing and blind answer recovery check the resulting annotations. To study the utility of this supervision, we develop MemSteer-Opt, which combines QA-based design evaluation with evidence-guided failure diagnosis to revise memory store, retrieval, and context construction without updating model parameters. Experiments on ATM-Bench and RHELM show that this supervision is competitive with official-QA-guided optimization. Relative to Base-BM25, MemSteer-Opt improves ATM-Bench Question Score by 13.4 points and RHELM accuracy by 4.1 and 6.6 points without and with external data, respectively. The supervision also benefits two existing optimizers on ATM-Bench, while new-user experiments demonstrate continued design optimization using supervision generated from the target histories.