Capability-Driven Self-Evolution of Agent Memory
Yaoqi Chen, Yuru Feng, Qi Chen and colleagues at Microsoft with USTC and UCSD introduce PrisMem, which evolves an executable agent-memory program by tracking progress on individual memory capabilities rather than only on the overall score.
Ask this paper
Motivation. In holistic memory evolution, 80.5% of revisions that do not raise the overall score still improve at least one capability. Judging only by the total discards those partial gains.
Method. Dependency-aware selection picks which capability to improve next, history-guided diagnosis refines a specialist program for that capability, and trace-guided integration compares programs on paired cases to merge complementary gains into one memory program.
Results. Evolved on shorter splits and tested on million-token histories, PrisMem beats the strongest baselines by 10.54 pp on BEAM-1M and 7.83 pp on LongMemEval-M. With GPT-5.5 it leads by 6.17 to 20.41 pp on BEAM-1M.
Ablations. Removing capability-driven evolution costs 10.15 pp, random diagnostic-case selection costs 5.26 pp, and dropping trace-guided integration costs 3.16 pp. Compact diagnostic contexts cut case tokens by 43.8% and evolution cost by 4% to 30%.
Abstract
Memory self-evolution uses task feedback to iteratively improve executable memory programs that store and retrieve information from past interactions. Existing approaches typically adopt holistic evolution, deriving revision directions from mixed feedback and judging progress by overall performance. This can obscure optimization directions and hide capability-specific gains offset by regressions elsewhere, leaving promising directions underexplored. We introduce capability-driven evolution, which extends search guidance from overall performance to individual capability dimensions, preserving promising revisions and expanding exploration beyond the boundaries of holistic evolution. We propose PrisMem, which uses dependency-aware capability selection to prioritize targets with potential cross-capability benefits and history-guided diagnosis to refine capability specialists. Trace-guided integration compares evaluated programs on paired differential cases, using their behavioral differences to consolidate complementary gains into a unified memory program. Experiments show that PrisMem outperforms the strongest baselines by 10.54 and 7.83 percentage points on BEAM-1M and LongMemEval-M, respectively, demonstrating its effectiveness on million-token histories.