在日常软件开发中使用大语言模型进行编码时，一个关键的效率问题正在浮现。[1]开发者借助LLM完成代码修改任务需要投入大量工作来建立充分的上下文，但这些在单个会话中生成的知识在实际应用中大部分会丧失。[1]这种上下文的流失导致多个层面的负面影响：团队成员在使用不同LLM模型（如Codex与Claude）时会反复消耗token来处理相同的信息，[1]开发者在重新启动会话或提交代码时需要重新加载已经处理过的上下文，造成token的浪费。[1]
这一问题的紧迫性日益凸显。[1]大型代码变更因缺乏充分的文档和上下文记录，使代码审查的难度大幅增加。[1]当前LLM模型间缺乏可移植性，不同的AI助手无法接续工作，进一步削弱了团队协作的效率。[1]业界因此呼吁解决这一根本问题，使不同的AI模型能够共享和传递开发过程中积累的知识，从而消除重复劳动，提升整体的开发效率。[1]
The use of large language models as coding agents in software development introduces a significant efficiency problem: the contextual knowledge these models generate during a single session is largely lost once the conversation ends [1]. This ephemeral nature of LLM-generated context creates cascading inefficiencies in development workflows. When developers restart sessions or hand off work to teammates, they must reconstruct and reload context that has already been processed, resulting in wasteful token consumption and repeated computational effort [1].
The lack of portable AI knowledge compounds collaboration challenges across teams. Different team members using different LLM models—such as Codex versus Claude—leads to substantial token duplication as each model must rebuild the same contextual understanding [1]. Additionally, large code changes that lack sufficient documentation and context records become significantly harder to review, as crucial reasoning and decision-making artifacts are not retained or transferred between review stages [1]. The author identifies LLM portability as an urgent problem requiring attention to enable AI models to seamlessly continue work across different platforms and systems [1].