references: Open Claw vs Hermes Agent: Which AI Agent Should You Use?

This video is talking about OpenClaw and Hermes Agent and to see which one fit you the best and if one can replace the other.

First of all, It's hard to say which one is objectively better than the other. This is kind of like design philosophy thing. Because if you see what each harness is trying to be from the very beginning, a lot of the behavior starts to make sense.

And once I saw that, it helps me to set up my own setup with the two agents. Let's start with Open Claw.

Open Claw is officially describes itself as a self-hosted gataway for agents across chat apps and channel surfaces.

Open Claw 官方将自己描述为跨聊天应用程序和渠道界面的代理的自托管网关。

And the docs are very explicit about that. You run one gateway process and the gateway becomes the bridge between your messaging apps and your agent and the gataway is the single souce of the truth of sessions, routing and channel connections.

文档对此非常明确。您运行一个网关进程，网关成为您的消息传递应用程序和代理之间的桥梁，而网关是会话、路由和通道连接的真实情况的单一来源。

And Hermes Agent is build by Nomic Research the docs describe it very differently. It calls itself a self-improving agent, not a wrapper, not a chatbox, not a coding copilot.

It actually describes itself as a self-improving agent. And the core idea is that it has a built-in learning loop.

它实际上将自己描述为一个自我改进的代理。核心思想是它有一个内置的学习循环。

It creates skills from the experience, improves them during use, nudges itself to like persist knowledge and build a deeper model of who you are accross sessions.

它从经验中创造技能，在使用过程中改进它们，推动自己喜欢持久的知识，并在整个会话中建立一个更深入的模型来了解你是谁。

So, right away you can see that Open Claw says, "Hey I'm a gateway." And Hermes Agent says, "I'm an agent that grows".

This is like the first big differences from their design philosophy. And to summarize that in one sentence, Open Claw is gateway first and Hermes is an agent loop first.

Open Claw feels like an AI communication and an automatic backbone. Well, Hermes fell like an AI worker that keeps learning. That's the cleanest comparison I can find. And once I start thinking about them that way, everything else start to put into places and make sense.

Open Claw 感觉就像是人工智能通信和自动主干。嗯，赫尔墨斯就像一个不断学习的人工智能工作者。这是我能找到的最清晰的比较。一旦我开始这样思考，其他一切就开始变得有意义了

Now, to be fair, they share a lot of commonality. That means you don't need to build one new framework for every workflow. Both of them have tools, have memories, have like skills and they have messaging surface and scheduled automation. So this is not a story of completely opposites. They are much more closer than we think.

现在，公平地说，他们有很多共同点。这意味着您不需要为每个工作流程构建一个新框架。他们都有工具，有记忆，有类似的技能，他们有消息传递界面和预定的自动化。所以这并不是一个完全相反的故事。他们比我们想象的要近得多。

But they diverge in where they put the center of gravity. And for me, the biggest difference is that memory versus skill growth.

但他们在重心的放置位置上存在分歧。对我来说，最大的区别是记忆力与技能的增长。

Open Claw's memory system is very explicit and it's very visible, very file system oriented. It's documents describe memory.md as like long-term memory, daily note files under memory. a certain date of time.md and they have an optional dreams.md for background consolidation and promotion into durable memory.

Open Claw 的内存系统非常明确，非常明显，非常面向文件系统。它的文档将 memory.md 描述为就像长期记忆，内存下的日常笔记文件。某个日期的 time.md 并且他们有一个可选的 dream.md 用于背景整合和提升到持久内存中。

Of course, their design may subject to change as they keep evolving this really, really fast, but that is the current structure and that matters. Because Open Claw's memory does not feel hidden. It feels like a alot of files, like a notebook, a long-term notebook, a daily notebook, a reflection notebook.

当然，随着他们不断快速发展，他们的设计可能会发生变化，但这就是当前的结构，这一点很重要。因为张开爪的记忆并不感觉被隐藏。感觉文件很多，像笔记本，长期笔记本，日常笔记本，反思笔记本。

It feels like a system that wants to know you better over time, wants to know you better over time. Hermes also has memory and I want to be careful here because I don't think it's fair o say Hermes does not care about memory. It absolutely does care about memory. The docs says that it builds a deepening model of who you are across sessions and it has persistent memory that grows over time.

感觉这是一个随着时间推移想要更了解你的系统，想要随着时间推移更了解你。Hermes 也有记忆功能，我在此要谨慎表述，因为我不认为说 Hermes 不重视记忆是公平的。它绝对重视记忆。文档中提到，它会在不同会话之间构建一个关于你身份的日益深入的模型，并且拥有随时间推移而不断增长的持久化记忆。

But the feel is different. Open Claw memory feels like, "Let me remember you." And Hermes memory feels like, "Let me improve my ability" to work with you after I know you." And that's the subtle difference, but it's kind of differentiate them from the design philosophy.

但两者的感觉不同。“Open Claw”的记忆给人一种“让我记住你”的感觉，而“Hermes”的记忆则让人觉得“等我了解你之后，我要提升与你协作的能力”。这就是其中的微妙差异，但也正是这种差异从设计理念上将它们区分开来。

Now, here's where Hermes gets really interesting. Hermes treats skills like first-class citizens and not just at the prompt level. It's designed in the harness level.

现在，Hermes 的真正精彩之处就在这里。Hermes 将技能视为第一类对象，而且不仅限于提示符层面，这种设计理念在框架层面就已经确立。

It docs explicitly say that the agent creates skills after complex tasks, especially after five tools calls. And when it discovers a very useful workflow, it recovers from mistakes or gets corrected by the user, it starts to build these skills. And that is a huge design choice

文档中明确指出，代理会在执行复杂任务后，尤其是调用五次工具后，开始构建技能。当它发现一个非常有用的工作流，或者从错误中恢复过来，或是被用户纠正时，它就会开始构建这些技能。这是一个重大的设计决策。

And that is a huge design choice because it means Hermes is not just hoping the model becomes proactive. It actually wiring skill growth into the loop. And that's a very strong opinionated design.

这是一个重大的设计决策，因为这意味着爱马仕不仅仅希望模型能够主动出击，而是真正将技能提升机制融入了系统循环之中。这是一种立场非常鲜明的设计。

Now, Open Claw absolutely has skills too. It has a nice architecture around the skills. Ans tools are callable functions, skill are marked like markdown instructions that teach the agent how and when to use these tools.

当然，Open Claw 也具备技能。它的技能架构设计得相当出色。工具被定义为可调用的函数，而技能则像 Markdown 指令一样进行标记，用于指导Agent在何时以及如何使用这些工具。

And plug-in plug-ins at channels and capabilities around that. So, it is a very elegant system.

此外，还可以通过插件扩展相关通道和功能。因此，这是一个非常优雅的系统。

But Open Claw does not feel as hardwired around the skill growth as Hermes does. Open Claw feels more like, "Hey, here are the tools and this is the workspace, the memory and the channels. Now, go be smart."

但与Hermes相比，Open Claw 在技能提升方面并不显得那么刻板。Open Claw 更像是这样说的：“嘿，这里有工具，还有工作区、内存和通道。现在，就靠你自己发挥智慧吧。”

Hermes feels more like, "No, no, no, no, it's not that. This is the workflow keeps happening, formalize it. Formalize the workflow, formalize the procedure. So, next time we run into it, make less mistakes and make this become more formal and more efficient"

Hermes给人的感觉更像是：“不，不，不，不，不是那样的。这是不断重复的工作流程，把它规范化。规范化工作流程，规范化操作程序。这样，下次遇到这种情况时，就能减少失误，让整个流程变得更加规范、更高效。”

And that differences show up in behavior.

This is my personal metaphor, not the in the official docs, but it's the one that helps me to understand both products.

这是我个人的比喻，不是官方文档中的比喻，但它可以帮助我理解这两种产品。

So, to me, Open Claw feels like a family member because it knows my like history knows my habits, it remembers little things or facts about me. It has more that relationship energy. Hermes feels like an experienced employee. It learns the workflows it gets sharper at repeated tasks.

所以，对我来说，Open Claw 感觉就像一个家庭成员，因为它了解我，就像历史了解我的习惯一样，它记得关于我的小事或事实。它具有更多的关系能量。Hermes给人的感觉就像是一位经验丰富的员工。它学习工作流程，在重复执行任务时变得更加清晰。

It accumulates produral knowledge and becomes more effective at doing that job. And that's the deepest the differences I feel and when I use them.

它积累了生产知识，并且在完成这项工作时变得更加有效。这是我在使用它们时感受到的最深刻的差异。

Now, about the setup. If you already run Open Claw because it's the pioneer and it came first of its kind as I just said, adding Hermes is actually not that hard. You can install Hermes directly the normal way. Nowadays, it's just like one line of command line of many of these tools. But Hermes also has an Open Claw migration path that can import compatible memories. So, the MD command allow list, user skills and even configure even share the configuration of the like the model configurations.

现在，关于设置。如果你已经运行了 Open Claw，因为它是先驱，而且正如我刚才所说，它是同类产品中的第一个，那么添加 Hermes 实际上并不难。您可以按照正常方式直接安装Hermes。如今，它就像许多这些工具的一行命令行。但 Hermes 还提供了一条 Open Claw 迁移路径，可以导入兼容的内存。因此，MD 命令允许列表、用户技能，甚至配置，都可以共享类似于模型配置的配置。

So, if you already have Open Claw your machine and in my case, it's a Mac mini, Hermes does not feel like starting from zero. You can literally ask your open Claw to install the Hermes for you and configure as much as you have that in Open Claw and even migrate some of the existing like memories or skills to Hermes. So, you're not absolutely not starting from zero.

因此，如果您的机器已经拥有 Open Claw，并且就我而言，它是 Mac mini，Hermes 不会觉得从零开始。您实际上可以要求您的 Open Claw 为您安装 Hermes，并根据您在 Open Claw 中的配置进行尽可能多的配置，甚至将一些现有的记忆或技能迁移到 Hermes。所以，你并不是绝对不是从零开始。

It lowers the friction a lot. So, what I like immediately in Hermes, the first thing I like in Hermes is the transparency. I like seeing more clearly what it's doing, What commands it is running and how it is progressing, what it was touching.

这大大降低了操作的阻力。因此，我最先喜欢上 Hermes 的地方，就是它的透明度。我喜欢能更清楚地看到它在做什么、正在执行哪些命令、进展如何，以及它访问了哪些资源。

And Hermes explicitly support streaming tool output and verbose models. And I feel that in practice, it gives me more confidence. The second thing I like was interruption. I actually being able to like cut in and interrupt what it is doing if I made a mistake in describing my task in my like in the previous message. And you can cut in and interrupt the current execution.

而且 Hermes 明确支持流式工具输出和详细模式。 我觉得在实际操作中，这让我更有信心。我喜欢的第二点是中断功能。如果我在之前的消息中描述任务时出了错，我实际上可以随时插手并中断当前的执行。你可以随时插手并中断当前的执行。

I don't always want to wait for a long chain of like fully finished before staring it to something else. And in my experience, Hermes feel better for that regard. For repeated workflows, Hermes also felt more focused, expecially on like operational routines.

我并不总是愿意等到一长串任务全部完成后才开始做其他事情。根据我的经验，在这方面 Hermes 的体验更好。对于重复性的工作流程，Hermes 也显得更加专注，特别是在处理日常操作时。

And in my case, I tried it on like trading related workflows. Not because Hermes is the like the trading expert, but because the design around the procedures and skill growth make me trust it more on repeated works, on learning the past experience and learning the past mistakes.

就我个人而言，我尝试将其应用于交易相关的工作流程。这并非因为Hermes是交易领域的专家，而是因为其围绕流程和技能提升的设计，让我在处理重复性工作时更加信赖它，它能帮助我总结过往经验并吸取过去的教训。

But Open Claw still feels richer as like a broader assistant. It has a better tool support for like browser for browser operations.

但 Open Claw 仍然感觉更丰富，就像一个更广泛的助手。它对浏览器操作有更好的类似浏览器的工具支持。

One thing I noticed is that Open Claw has a better managed the browser capability. Meaning it can launch the managed browser as I can like type in the credentials and it can continue to do tasks. But I have a hard time to do that with my Hermes Agent. Maybe it was just me, but it feels like Open Claw still has like a richer and a broader assistant.

我注意到的一件事是 Open Claw 可以更好地管理浏览器功能。 这意味着它可以启动托管浏览器，因为我可以输入凭据，并且它可以继续执行任务。但我和我的Hermes Agent很难做到这一点。也许只有我这么认为，但感觉 Open Claw 仍然有一个更丰富、更广泛的助手。

And I think the documents explains why. It is built around gateway, routing all the tools, workspaces, channels, plug-ins and explicit memory structure.

我认为文档中解释了原因。该系统以网关为核心，负责路由所有工具、工作区、频道、插件以及明确的内存结构。

So, when I use Open Claw, I feel more of this. This thin knows me and this thing has is resourceful, it has a lot of capability to invoke the tools and to leverage the plugins. This thing spans my surfaces. This thing has a broader sense of context.

所以，当我使用Open Claw时，我更有这种感觉。这个瘦子认识我，而且这个东西足智多谋，它有很多能力来调用工具和利用插件。这东西横跨我的表面。这件事有更广泛的背景意义。

It feels lees narrowly procedural and more like a long-term assistant.

它给人的感觉是狭隘的程序性，更像是一个长期的助手。

About security, both systems take security seriously, but they expose it differently.

关于安全性，这两个系统都非常重视安全性，但它们的暴露方式不同。

Hermes documents a defense-in-depth model, and its messaging documents says that the gateway denies users by default unless they're allowed to pair with via the end.

Hermes 采用深度防御模型，其说明文档指出，网关默认会拒绝用户，除非用户已获准通过终端进行配对。

Open Claw, meanwhiles, has very explicit controls around a lawless mansion rules, EM, and group policies and routing.

与此同时，Open Claw 在无序庄园规则、EM、组策略和路由方面都设有非常明确的控制措施。

So, I would not frame this as like one is secure and the other one is careless, but both have paid a lot of attentions to the securities, and which is very important, and it's essential.

因此，我不会将这一点描述为一个是安全的，另一个是粗心的，但两者都对安全给予了很多关注，这是非常重要的，也是必要的。

Here's the question I started with: Is one replacing the other? And for me, the answer is now no.

这是我一开始提出的问题：一个会取代另一个吗？对我来说，现在的答案是否定的。

At the beginning , I thought I had to pick one because like Hermes agent is going so fast, I thought I'm going to like use Hermes as my AI main agent to take replace my Open Claw setup, but now I'm happily let them both coexist because they're not actually trying to be the same thing, and Open Claw is one I trust more as the always there assistant, and Hermes is the one I increasingly trust for certain workflow executor that keeps getting better and better, and that's a very different pairing.

起初，我以为必须二选一，因为像Hermes这样的AI Agent发展得太快了，我本打算把它作为我的主要AI Agent来取代Open Claw的配置，但现在我很乐意让它们共存，因为它们其实并不试图成为同一类工具， Open Claw 是我更信赖的常驻助手，而 Hermes 则是我越来越信赖的特定工作流执行者，它正变得越来越好，这二者搭配起来截然不同。

And once I stopped forcing them into a a like one replacing the other, one winner and one loser frame, the whole thing makes more sense to me, and it becomes easier.

一旦我不再强行将它们套入“一个取代另一个”、“一个赢家一个输家”的框架中，整件事对我来说就变得更有道理了，也变得更容易了。

So if you want to shortest possible conclusion, choose Open Claw if you want one assistant across many surfaces, richer routing and gateway behavior, explicit work workspace memory, have a better use of like manage the browsers and the variety of tools that the community has built up so far, and more of that long-term like personal assistant feel that knows you and gets to know your better

因此，如果您想要最短的结论，请选择 Open Claw，如果您想要一个跨多个表面的助手、更丰富的路由和网关行为、明确的工作工作区内存、更好地使用管理浏览器和社区迄今为止建立的各种工具，以及更多像个人助理一样的长期感觉，了解您并更好地了解您.

And you choose Hermes if you want a like self-improving autonomous agent. This is the core design principle of Hermes.

如果你想要一个类似于自我提升的自主代理，那就选择 Hermes。这就是 Hermes 的核心设计原则。

A stronger emphasis on the procedural skill growth. This is also very important cuz it learns and grows the skill over time when you use it.

更加注重程序技能的提升。这一点也非常重要，因为随着时间的推移，当你不断使用它时，技能就会得到学习和提升。

Ans if you need a runtime that wants to live everywhere, and sharper at repeated workflows, Hermes may be the better choice. You can also do something like me.

如果需要一个能无处不在、且在处理重复性工作流时更高效的运行时，Hermes 可能是更好的选择。你也可以像我一样做些类似的事情。

You may end up wanting both because one feels like a family member and a resourceful tool user, and other feels like a very good employee.

您最终可能会两者都想要，因为一个感觉像家庭成员和足智多谋的工具用户，而另一个感觉像非常好的员工。

And honestly, that's a pretty good combo. If you've been using either Open Claw or Hermes, I'd love to know what role each one you play in your setup. Are you trying to replace one with the other, or are you like starting to see the same thing I'm seeing? Like they're actually better together in different scenarios.

老实说，这是一个非常好的组合。如果您一直在使用 Open Claw 或 Hermes，我很想知道您在设置中扮演什么角色。你是想用另一种来代替一种，还是你想开始看到我所看到的同样的东西？就像他们在不同的场景中实际上在一起表现得更好一样。