📄 中文摘要
许多构建 AI 代理的团队在使用大规模语言模型（LLM）时遇到一个常见问题：即使在拥有 200,000 令牌的上下文窗口并加载了整个知识库后，代理在对话中仍会“忘记”关键信息。问题并不在于上下文窗口的大小，而在于使用方式。现代 LLM 拥有令人印象深刻的上下文窗口，例如 GPT-5.2 支持 400,000 令牌，Claude Sonnet 4.6 支持 200,000 令牌，但如何有效利用这些上下文窗口是关键。
上下文窗口的谎言：为什么你的 AI 代理会忘记一切
出处: The Context Window Lie: Why Your AI Agent Forgets Everything
发布: 2026年3月11日
📄 English Summary
The Context Window Lie: Why Your AI Agent Forgets Everything
Many teams building AI agents encounter a common issue when using large language models (LLMs): even with a 200,000 token context window and a fully loaded knowledge base, the agent still 'forgets' critical information during conversations. The problem lies not in the size of the context window, but in how it is utilized. Modern LLMs boast impressive context windows, such as GPT-5.2 with 400,000 tokens and Claude Sonnet 4.6 with 200,000 tokens, yet the effective use of these context windows is crucial.