一句话：借助人类读代码的EEG信号蒸馏事件边界控制器，为仓库级代码智能体构建紧凑情景记忆以减少检索抖动。
仓库级代码智能体在导航与修改大型代码库时，其记忆机制往往十分脆弱：扁平的对话历史或向量存储会导致检索过载、重复查询与上下文不一致。本文利用人类阅读代码时采集的脑电（EEG）信号，依据事件分割理论（Event Segmentation Theory）标注事件边界，并据此蒸馏出一个边界控制器。该控制器在运行时仅依赖软件工程信号，即可写出紧凑的“情节卡片”（Episode Cards），从而减少检索抖动并提升智能体记忆的稳健性。摘要未给出具体数据集、基准或量化结果，仅以缓解检索抖动、增强稳健性为评估目标。
Code agents are increasingly used to navigate and modify large software repositories, but their agentic memory is often fragile: flat chat histories or vector stores can cause retrieval overload, repeated re-querying, and inconsistent context. We leverage EEG (electroen-cephalography) during human code reading to label event boundaries predicted by Event Segmentation Theory. We propose a distilled boundary controller that relies only on software-engineering signals at runtime to write compact Episode Cards, reducing thrashing and improving robustness.