📄 中文摘要
LLM（大语言模型）在对话中能够记住用户的名字，但在大约20-30条消息后，之前的上下文将被遗忘。传统的解决方案是使用向量搜索，通过嵌入对话并在后续检索相关信息。然而，这种方法在面对相互矛盾的事实、关键信息与琐碎信息同样衰减，或系统在用户提及不同药物时悄悄覆盖药物过敏信息时，效果不佳。widemem是一个开源的记忆层，能够处理向量搜索无法解决的部分问题，提供批量冲突解决功能，确保信息的准确性和一致性。
为您的本地 LLM 提供真正有效的记忆
出处: Give Your Local LLM a Memory That Actually Works
发布: 2026年3月11日
📄 English Summary
Give Your Local LLM a Memory That Actually Works
LLMs (Large Language Models) can remember a user's name but tend to forget everything else after about 20-30 messages, as the context window fills up. The typical solution involves vector search, embedding conversations, and retrieving relevant chunks later. However, this approach struggles when faced with contradictory facts, critical information decays at the same rate as small talk, or when the system quietly overwrites important details like drug allergies due to mentions of different medications. Widemem is an open-source memory layer designed to address the limitations of vector search, offering batch conflict resolution to ensure the accuracy and consistency of information.