📄 中文摘要
开发者通常将上下文窗口视为数据库，但实际上它更像是一个工作桌。代理在开始任务时，桌子上所放置的内容就是它对该任务的全部知识。之前的运行、学习的行为和先前的决策都不存在，除非在代理开始之前将这些信息放在桌子上。这一区别改变了调试代理的方式。当代理出现意外行为时，开发者往往会将责任归咎于模型，认为是模型产生了幻觉或编造了内容。然而，实际上，代理通常只是根据任务开始时可用的信息做出了合理的反应，问题往往在于上下文不完整、过时或缺少正确的锚点。
上下文窗口是桌子，而不是记忆
出处: The Context Window Is a Desk, Not a Memory
发布: 2026年3月9日
📄 English Summary
The Context Window Is a Desk, Not a Memory
Most developers treat the context window as a database, but it functions more like a working desk. The information available on the desk when an agent starts a task constitutes its entire knowledge for that task. Previous runs, learned behaviors, and prior decisions do not exist unless deliberately placed on the desk beforehand. This distinction significantly alters the approach to debugging agents. When an agent behaves unexpectedly, the instinct is to blame the model, claiming it 'hallucinated' or 'made something up.' However, more often than not, the agent acts reasonably based on the information available at the start of the task, with issues arising from incomplete, outdated, or missing context.