知识/记忆管理开放协议
1 分•作者: iamalnewkirk•大约 1 个月前
最近，我开始思考将我的 AllSign 项目的核心基础开源会是什么样子。随着我越陷越深地钻研这个概念，我意识到这些核心基础实际上很好地解决了我之前遇到的一个关于代理记忆的问题。
我读了一篇题为“基于文件系统的 LLM 代理记忆”的论文，链接是：https://arxiv.org/pdf/2607.26637。该论文（正确地）得出结论：“信息收集和组织并不能自动带来智能”。我对此表示赞同。
我心想，缺失的是“策展”。必须有人来决定什么值得保留，什么可以被替换，什么才是真实的，以及什么悄无声息地不再真实。
需要支持 HITL（人工干预）和 HOTL（机器干预）的知识/记忆管理。
然后我想到，如果能将 AllSign 的核心基础变成一个支持任何参与者的知识管理开放协议呢？
这促使我创建了 Facts 协议和 Fact CLI。
如果您有兴趣，可以在这里阅读更多信息：https://gist.github.com/iamalnewkirk/22ae2d8c8cc1b8790236da9edf2f7b73
- https://github.com/facts-kms/cli
- https://github.com/facts-kms/spec
查看原文
Recently, I got to thinking about what it might look like to open-source the core fundamentals of my AllSign project. As I fell further and further down that conceptual rabbit hole, it occurred to me that those core fundamentals are actually a really good solution to a problem I saw with agent memory.<p>I read this paper called "Filesystem-Based Memory for LLM Agents", https://arxiv.org/pdf/2607.26637, which concluded (rightfully so) that "information gathering and organization does not automatically get you intelligence". I agree.<p>I thought to myself, what's missing is "curation". Something has to decide what's worth keeping, what replaces what, what's true, and what quietly stopped being true.<p>Knowledge/memory management that supports HITL and HOTL.<p>Then I thought, what if those AllSign core fundamentals could be turned into an open protocol for actor-agnostic knowledge management?<p>That led me to create the Facts protocol and Fact CLI.<p>If you're curious, you can read more here: https://gist.github.com/iamalnewkirk/22ae2d8c8cc1b8790236da9edf2f7b73<p>- https://github.com/facts-kms/cli
- https://github.com/facts-kms/spec