Cohub：面向 Agent 的分层持久化工作空间与记忆架构
this new open-source github repo gives agents a persistent workspace beyond the…
this new open-source github repo gives agents a persistent workspace beyond the conversation, built around an agent harness.
这个新的开源 GitHub 仓库为智能体提供了超越对话的持久化工作空间，其构建围绕一个智能体框架（agent harness）展开。
Cohub is an open-source agent OS that gives agents one shared place for files, sessions, tools, scheduled tasks, and reusable context.
Cohub 是一个开源的智能体操作系统，它为智能体提供了一个共享的文件、会话、工具、定时任务和可复用上下文的统一场所。
I like the 2-layer memory model Cohub ended up with.
我喜欢 Cohub 最终采用的双层记忆模型。
The context hub is curated organizational memory. A Space is the working state of one project.
上下文中心是经过策展的组织记忆。Space 是单个项目的工作状态。
Those need very different write policies.
这两者需要截然不同的写入策略。
If every temporary script, draft, failed attempt, and personal note enters global memory, retrieval quality eventually collapses. But deleting all of that after a task finishes is wasteful too, because the failed attempts and intermediate artifacts often explain how the result was reached.
如果每个临时脚本、草稿、失败的尝试和个人笔记都进入全局记忆，检索质量最终会崩溃。但任务结束后删除所有这些内容也是浪费，因为失败的尝试和中间产物通常解释了结果是如何达成的。
So Cohub keeps the full work-site inside the Space and promotes only durable knowledge into the Hub.
因此，Cohub 将完整的工作现场保留在 Space 内部，仅将持久的知识提升到 Hub。
That gives the harness two retention horizons: preserve execution state cheaply, curate shared judgment selectively.
这为框架提供了两种保留期限：低成本地保留执行状态，有选择地策展共享判断。
A lot of agent memory systems seem to jump straight from conversation history to "long-term memory." This is a more useful distinction.
许多智能体记忆系统似乎直接从对话历史跳转到“长期记忆”。这是一个更有用的区分。
An organization probably does not need one giant memory. It needs somewhere agents can make a mess, and somewhere they cannot.
组织可能不需要一个巨大的单一记忆。它需要的是智能体可以随意搞乱的地方，以及它们不能随意搞乱的地方。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力