当前AI Agent采用的记忆插件架构存在多个关键缺陷。1这些插件通过处理会话记录生成记忆片段，将其插入向量数据库，再在每次提示时检索最相似的片段，但这一流程面临相似度检索精度低、片段存储丧失上下文、过去记录不可靠、Agent无法识别知识空缺等问题。1此外，这类存储库难以审计，对系统的可追溯性和可控性构成挑战。1
针对这些问题，有开发者提出了以Markdown文档为基础的替代方案。1该方案无需向量数据库、嵌入或后台守护进程，所有内容均为可读写的Markdown文档，Agent在工作前查阅相关文档，工作后更新相关内容，将传统的"提示→构建→遗忘"流程改为"提示→查阅→构建→更新"。1开发者开源了名为Operator Memory的插件来实现这一方案，该方案源自其一年前开始使用的internal文件夹方案，用于让Agent记录规范、计划和索引。1
An author has challenged the conventional approach to artificial intelligence agent memory systems, arguing that traditional memory plugins are fundamentally flawed and proposing a documentation-based alternative instead.1 The current architecture for agent memory typically processes conversation records, generates memory fragments, inserts them into a retrieval-augmented generation (RAG) database, and retrieves the five most similar fragments during each prompt—yet this approach suffers from significant limitations.1 Specifically, similarity-based retrieval lacks precision, stored fragments lose their original context, historical records become unreliable, agents cannot identify knowledge gaps in their own understanding, and storage repositories become difficult to audit.1
Rather than relying on vector databases and embedding systems, the author has been using an internal folder structure for over a year to enable agents to record standardized procedures, plans, and indexes in human-readable formats.1 This led to the development of the Operator Memory plugin, which implements a documentation-based memory system using only readable and writable Markdown files—eliminating the need for vector databases, embeddings, or background processes.1 The workflow has been transformed from a linear sequence of "prompt, build, forget" to a cycle of "prompt, consult, build, update," allowing agents to review relevant documentation before working and update records after completion.1 The plugin has been released as open source.1
评论
还没有评论，欢迎留下第一条。