📄 中文摘要
持久记忆不仅仅是为大型语言模型（LLM）存储笔记，它还会影响LLM对问题的思考方式。在相同的模型、相同的提示和相同的温度下，不同的记忆支架会产生结构上不同的解决方案。研究团队在开发环境中积累了640多个持久记忆，这些记忆来源于数百次Claude Code会话。实验结果表明，记忆的构建方式对生成的结果有显著影响，验证了持久上下文在AI推理中的重要性。
记忆支架如何影响大型语言模型推理：持久上下文如何改变人工智能的构建
📄 English Summary
Memory Scaffolding Shapes LLM Inference: How Persistent Context Changes What AI Builds
Persistent memory serves not only as a storage for notes in large language models (LLMs) but also significantly influences how these models approach problem-solving. The same model, prompt, and temperature can yield architecturally distinct solutions when different memory scaffolding is applied. The research team accumulated over 640 persistent memories from hundreds of Claude Code sessions in a development environment. The experimental results demonstrate a significant impact of memory construction on the generated outcomes, validating the importance of persistent context in AI inference.