The paper introduces MACE, a framework that co-evolves memory organization and use through execution feedback. Empirical studies show that grouping dependencies into functional memory units improves retention and connecting units increases retrieval of jointly needed units and links. MACE uses MemGoG, a graph of subgraphs representing functional units, and a loop that selects task-relevant units, presents them as instructions or checklists, and updates scores and relations based on outcomes.
On eight benchmarks, MACE achieved an average score of 81.11%, outperforming the strongest baseline SAGE at 78.97%. The framework also demonstrates that updating choices from the paired outcomes of a combination and its presentation format outperforms scoring combinations and formats separately. This work is interesting because it highlights how adaptive memory organization and presentation can improve multi-agent system performance.