MemSteer converts unlabeled user histories into evidence‑grounded question‑answer (QA) supervision to improve an agent's memory system. It first builds an answer plan that specifies how facts from related records jointly support a reference answer; this plan guides question generation. Source‑grounded auditing and blind answer recovery are then used to verify the quality of the generated annotations.
Using this supervision, MemSteer‑Opt couples QA‑based design evaluation with evidence‑guided failure diagnosis to revise the memory store, retrieval, and context construction without touching model parameters. Experiments on ATM‑Bench and RHELM show that MemSteer‑Opt outperforms a BM25 baseline by 13.4 points on the ATM‑Bench question score and improves RHELM accuracy by 4.1 and 6.6 points with and without external data, respectively. The supervision also benefits existing optimizers and, in new‑user tests, demonstrates that continual design optimization is possible by generating supervision from target histories.
Review