Splitting agent memory into a planning bank and an execution bank, and attributing blame before writing, beats outcome-only memory

arXiv 2609.02074·medium signalXBlueskyLinkedInCopy link

CHIME (arXiv 2609.02074, 2026-09-02) identifies a credit-assignment bug shared by every self-evolving memory scheme: they feed on final task outcomes, which conflate plan quality with execution errors and environmental noise, so the accumulated experience is biased. CHIME keeps separate planning and execution banks and follows an attribute-before-memorize rule, first assigning each outcome to the plan, the execution, both, or neither, then updating only that bank. It beat training-based and self-evolving baselines on four long-horizon benchmarks with far fewer stored items, found planning memories to be worth more than execution memories, and showed the accumulated memory transfers across backbone models. Code is promised at github.com/ATH-MaaS/Marco-DeepResearch (331 stars, last pushed 2026-08-09).