Planning enables agents to decompose complex long-horizon tasks into manageable steps. Test-time search and training-based approaches improve planning but either raise inference cost or require costly training data. Self-evolving memory accumulates reusable experience in an external bank, allowing planning ability to improve at inference without parameter updates. Existing self-evolving memory suffers from a credit assignment problem: it relies on final task outcomes as feedback, which conflate plan quality, execution errors, and environmental factors, resulting in biased and noisy memories. To solve this, we introduce Credit-Aware Hierarchical Memory Evolution (CHIME). CHIME keeps separate planning and execution banks and follows an attribute‑before‑memorize principle: it first attributes each outcome to the plan, the execution, both, or neither, then updates only the relevant bank. Extensive experiments on four long-horizon benchmarks show CHIME consistently outperforms state-of-the-art training‑based and self‑evolving memory baselines. Analyses reveal that CHIME builds effective memory with far fewer items, that memory values faithfully reflect downstream utility—high‑quality planning memories are more valuable than execution memories—and that the accumulated memory transfers across backbone models. Code will be released at https://github.com/ATH-MaaS/Marco-DeepResearch.
Review