跳到正文
@rohanpaul_ai· @rohanpaul_ai · X·· 2026-09-02AI 评分52
AI 导读

开源项目 Cohub 为智能体提供对话之外的持久工作区,把文件、会话、工具、定时任务和可复用上下文集中在一处。其记忆分两层,Space 保存单个项目的完整工作状态,Hub 只沉淀经过整理的长期知识,避免临时脚本、草稿和失败尝试污染全局记忆。项目 README 称其公司内部每周在 Cohub Spaces 中消耗 1000 亿 token。

正文

this new open-source github repo gives agents a persistent workspace beyond the conversation, built around an agent harness.

Cohub is an open-source agent OS that gives agents one shared place for files, sessions, tools, scheduled tasks, and reusable context.

I like the 2-layer memory model Cohub ended up with.

The context hub is curated organizational memory. A Space is the working state of one project.

Those need very different write policies.

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.

So Cohub keeps the full work-site inside the Space and promotes only durable knowledge into the Hub.

That gives the harness two retention horizons: preserve execution state cheaply, curate shared judgment selectively.

A lot of agent memory systems seem to jump straight from conversation history to "long-term memory." This is a more useful distinction.

An organization probably does not need one giant memory. It needs somewhere agents can make a mess, and somewhere they cannot.

来源:@rohanpaul_ai · x.com