Capability-Driven Self-Evolution of Agent Memory

Capability-Driven Self-Evolution of Agent Memory. By Yaoqi Chen and ten others, from the University of Science and Technology of China, Microsoft, and one other institution. Published on the fifth of October. No code repository is linked to it.

Memory self-evolution uses task feedback to improve, round after round, the executable programs that store and retrieve information from an agent's past interactions. Existing approaches usually evolve holistically. They derive revision directions from mixed feedback and judge progress by overall performance. That can obscure which way to optimize, and it can hide a gain in one capability that is offset by a regression somewhere else, leaving promising directions underexplored. A revision that makes the memory much better at one thing and slightly worse at another can look like no progress at all, and get thrown away.

The proposal is capability-driven evolution. Search guidance extends from overall performance down to individual capability dimensions. That preserves revisions that are promising on some dimension, and pushes exploration beyond the boundaries of holistic evolution. Progress is judged per capability, so a gain in one place is no longer hidden by a loss in another.

PrisMem puts this into practice with three components. Dependency-aware capability selection chooses which capability to target, prioritizing targets with potential benefits for other capabilities. History-guided diagnosis refines specialists for individual capabilities. Trace-guided integration then brings the specialists back together: it compares evaluated programs on paired differential cases, and uses their behavioral differences to consolidate complementary gains into one unified memory program.

PrisMem outperforms the strongest baselines by ten point five four percentage points on BEAM-1M, and by seven point eight three on LongMemEval-M, which the authors take as evidence of effectiveness on histories of a million tokens.

Here a single score is treated as the obstacle rather than the goal. Split the search into specialists that each improve one capability, then merge what they found, and gains that an overall average would have cancelled out survive long enough to be kept.