CHIME Enhances Agentic Planning with Credit-Aware Memory

Planning is a fundamental capability for AI agents, enabling them to break down complex, long-horizon tasks into manageable steps. While existing methods improve planning through test-time search or training, they often incur high inference costs or demand extensive training data. Self-evolving memory approaches, which accumulate reusable experience from agent interactions, offer an alternative by improving planning at inference time without parameter updates. However, these methods suffer from an inherent credit assignment problem, as they rely on final task outcomes that conflate plan quality with execution errors and environmental factors, leading to biased and noisy accumulated experience. To overcome this, researchers propose CHIME (Credit-Aware Hierarchical Memory Evolution), a novel self-evolving memory framework. CHIME addresses the credit assignment problem by maintaining distinct memory banks for planning and execution. It adheres to an "attribute-before-memorize" principle: each task outcome is first attributed to the plan, the execution, both, or neither, and only then is the corresponding memory bank updated. This targeted approach ensures that accumulated experience is more accurate and relevant. Extensive experiments across four long-horizon agent benchmarks demonstrate that CHIME consistently outperforms state-of-the-art training-based and self-evolving memory baselines. Further analysis reveals that CHIME accumulates effective memory with significantly fewer items, and the learned memory values accurately reflect downstream utility, with planning memories proving more valuable than execution memories. Additionally, the accumulated memory effectively transfers across different backbone models, highlighting CHIME's robustness and efficiency.