它不是又一个聊天机器人,也不是写代码的助手。它是一个「住在你电脑上、越用越聪明」的私人 agent。最近中文圈讨论很多,值得了解一下它到底是什么。
大多数 AI 工具是这样的:你打开一个对话,聊完关掉,下次再打开它什么都不记得了。Hermes Agent 不一样。它装在你自己的电脑或服务器上,一直在后台跑着。你上次让它做了什么、你的偏好是什么、它自己学会了什么技能——这些全都记着,下次直接用。
一句话版本:它是一个本地运行、持久记忆、自我进化的私人 AI agent。由 Nous Research 开发(就是做 Hermes 系列模型的那个实验室),MIT 开源,免费。
Hermes Agent 本身是命令行工具。最近小红书上讨论的是一个叫 Hermes Desktop 的项目——社区开发者 fathah 做的第三方桌面端,不是 Nous Research 官方出品,但把命令行里的操作搬到了图形界面里。
⚠️ Desktop 版是第三方社区项目,稳定性还没经过大规模验证。如果你要试,先从 Hermes Agent CLI 本体开始,确认好用了再加 Desktop 层。
| Hermes Agent | Claude Code | ChatGPT / Codex | |
|---|---|---|---|
| 定位 | always-on 私人 agent | session 制编程助手 | 通用 AI + 编程 |
| 运行方式 | 后台常驻 | 开 session 用完关 | 云端/本地 session |
| 记忆 | 三层,全本地 | CLAUDE.md + 自动记忆 | 云端记忆 |
| 自进化 | 自动写 skill 文件 | 不自动,你手写 | Custom GPTs |
| 消息触达 | Telegram/Discord/Slack/WhatsApp | 终端 + 手机遥控 | App + 手机 |
| 数据位置 | 你的机器 | 你的机器(代码),云端(对话) | 云端 |
| 上手门槛 | 一行命令安装,但配置多 | 一行命令安装 | 开浏览器就用 |
如果你想要一个「一直在后台帮你盯着事情」的 agent——汇总消息、定时跑任务、记住你的偏好、越用越懂你——Hermes 是目前最认真在做这件事的开源项目。
如果你只是需要一个帮你写代码、改文章、做分析的工具,Claude Code 和 ChatGPT 已经够了,没必要为了 Hermes 多维护一套系统。
目前它还在 v0.13,不是 1.0。每周发版说明迭代快,也说明还在频繁改动。想尝鲜的可以装,想稳定用的再等等。
不是因为 10万+ star。是因为它代表的方向:AI 从「你开一个对话用完关掉」变成「一个一直在你身边、越来越懂你的助手」。这个方向不管最后赢的是 Hermes 还是别的什么,都是 AI 工具接下来要走的路。
Claude 在做记忆和 Projects,ChatGPT 在做 Custom GPTs 和 Memory,Hermes 用开源和本地化走了最激进的一条路。三家路线不同,终点可能相似。
Hermes is not another chatbot or another coding helper. It is a private agent that lives on your machine and gets more useful as you use it.
Most AI tools work like this: you open a chat, finish the conversation, close it, and next time it remembers very little. Hermes Agent is built differently. It runs on your own computer or server, keeps working in the background, and remembers what you asked it to do, what you prefer, and what skills it learned along the way.
The short version: Hermes Agent is a local-first, persistent-memory, self-improving private AI agent. It is developed by Nous Research, the lab behind the Hermes model family, and released under MIT open source.
Hermes Agent itself is a command-line tool. The version being discussed recently is Hermes Desktop, a third-party desktop interface by the community developer fathah. It is not an official Nous Research product, but it brings many command-line operations into a graphical interface.
The Desktop version is a third-party community project and has not been validated at large scale. If you want to try Hermes, start with the core CLI first, then add the desktop layer if it proves useful.
| Hermes Agent | Claude Code | ChatGPT / Codex | |
|---|---|---|---|
| Positioning | always-on private agent | session-based coding assistant | general AI + coding |
| Runtime | background service | open a session, close it later | cloud/local sessions |
| Memory | three layers, local | CLAUDE.md + auto memory | cloud memory |
| Self-improvement | writes skill files | manual configuration | Custom GPTs |
| Messaging | Telegram/Discord/Slack/WhatsApp | terminal + mobile remote control | app + phone |
| Data location | your machine | your machine for code, cloud for chat | cloud |
| Learning curve | one-line install, many settings | one-line install | open browser and use |
If you want an agent that keeps watching things in the background - summarizing messages, running scheduled tasks, remembering preferences, and learning from repeated work - Hermes is one of the most serious open-source attempts right now.
If you only need help writing code, editing documents, or doing analysis, Claude Code and ChatGPT are already enough. There is no need to maintain another system just for the sake of it.
It is still at v0.13, not 1.0. Weekly releases show fast momentum, but also frequent change. It is good for experimenting, less ideal if you need a stable daily system today.
Not because of the 100K stars. Because it represents a direction: AI tools are moving from "open a chat, use it, close it" toward "an assistant that stays with you and gets to know your workflow." Whether Hermes wins or not, this direction matters.
Claude is building memory and Projects. ChatGPT has Custom GPTs and Memory. Hermes takes the most aggressive local-first and open-source route. The paths differ, but the endpoint may rhyme.