Hermes Agent

• 1. overview (704 chars)

  Hermes Agent is an open source self-evolving AI agent framework launched by Nous Research, an American AI research institution, on February 25, 2026. The project’s GitHub stars exceeded 70,000 in less than two months after its release, making it one of the fastest-growing projects in the global open source intelligent field.   Different from traditional AI tools, the core differentiation of Hermes Agent lies in its "self-evolution" ability - through the built-in learning closed-loop system, it can automatically accumulate experience and generate reusable skills after completing the task, achieving the effect of "getting smarter with more use". The project adopts the MIT license, supports multi-platform message access and 200+ large models, and is regarded by the industry as a representative product of the "AI 2.0" era.

• 2. Product background (1211 chars)

  Hermes Agent is developed by the Nous Research team. Nous Research was founded in 2023 with a team of about 20 people. The core members are from StabilityAI. It has raised a total of about US$70 million in financing, of which the 2025 Series A round was led by Paradigm, which raised US$50 million. The organization focuses on the research and development of open source AI Agent technology, and Hermes Agent is its core product.   From the perspective of product positioning, Hermes Agent aims to solve the "goldfish memory" problem of traditional AI assistants. Traditional AI loses all context after each conversation ends, and users cannot truly accumulate and reuse AI experience. Hermes Agent implements the two core capabilities of "persistent memory" and "automatic skill generation". After the user asks AI to complete a complex task, the system will automatically record the solution as a reusable skill document, which can be directly called when encountering a similar task next time without re-describing the requirements.   In April 2026, the Chinese AI team EvoMap publicly accused the Hermes Agent’s core self-evolution capability of systematically replicating its Evolver engine, triggering a heated discussion in the open source community on the ethics of innovation. Nous Research responded twice denying the allegations. This incident reflects the industry’s high concern for the originality of AI Agent technology.

• 3. Product functions and usage experience (1948 chars)

  Hermes Agent provides six core functions, corresponding to different usage scenarios.   The first is the persistent memory function. The system stores the user's project information, preferences, historical conversations and other data in the ~/.hermes/ directory. The longer it runs, the deeper it will understand the user. This means users don’t need to re-tell the AI ​​their project structure, coding style or workflow every time, the AI ​​will automatically remember and adapt.   The second is the automatic skill creation function. When a user asks Hermes Agent to solve a complex problem, the system will automatically generate a skill document and save it locally. These skills follow the agentskills.io open standard and can be called by other agents or shared to the skill library for community use.   The third is a multi-platform message gateway. Hermes Agent natively supports five international platforms: Telegram, Discord, Slack, WhatsApp, and Signal. At the same time, through the support of the Chinese community, it can also access domestic platforms such as WeChat, DingTalk, Feishu, and Enterprise WeChat. This allows users to interact with the AI ​​through their usual communication tools.   The fourth is scheduled automation tasks. The built-in cron scheduler supports setting up automated tasks such as daily report generation, nightly backups, and weekly audits. Users can pre-configure the time and frequency of task execution.   The fifth is parallel sub-agent. Users can generate multiple isolated sub-agents, each with independent dialogue windows and terminals, achieving zero context consumption through RPC. This is useful for scenarios where multiple tasks need to be handled simultaneously.   The sixth is complete browser and web page control. Built-in browser automation capabilities support web search, content extraction, visual analysis, image generation, text-to-speech and other functions, covering most daily development needs.   From the technical specifications, Hermes Agent supports the installation of Linux, macOS and WSL2 operating environments, and deployment can be completed with a curl command. In terms of model support, there is lock-free access to Nous Portal (OAuth method), OpenRouter (200+ models), as well as custom APIs and local vLLM.

• 4. Pricing and business model (600 chars)

  Hermes Agent adopts a completely open source business model, and the project adopts an MIT license, which means that anyone can use, modify and commercialize the product for free. Nous Research makes money mainly by providing paid model access services through its cloud service Nous Portal. Users can choose to use the open source version to self-host, or use official hosting services to obtain more stable API access.   From the perspective of target users, Hermes Agent is mainly aimed at developers and technical teams, especially enterprise users who need to run AI tasks for a long time and want to accumulate team knowledge. For individual developers, the free open source version can already meet most needs.

• 5. User reviews (1125 chars)

  Judging from feedback on social media and the developer community, Hermes Agent has received relatively positive user reviews.   Positive reviews mainly focus on several aspects. The first is to greatly improve efficiency. Some users reported that in the scenario of automatic deployment of projects, the time was reduced by more than 70% after using Hermes Agent, and the error rate was also significantly reduced. The second is real knowledge accumulation. Some users said that this is equivalent to a "technical colleague who will not leave". AI will remember all the team's technical decisions and project experience. In addition, there is multi-platform access convenience. Users can access AI at any time through common tools such as Discord or Slack without opening a specific web page or application.   Negative feedback mainly points to technical challenges. The first is the issue of API stability. Due to its high reliance on external large model services, network fluctuations will affect the user experience. The second is the complexity of configuration. Some users reported that the initial configuration and model access settings are relatively cumbersome for novices. In addition, Windows native support is still in the experimental stage, and the current running effect on Windows is not as stable as Linux and macOS.

• 6. Industry evaluation (1155 chars)

  From an industry perspective, the emergence of Hermes Agent is regarded as an important breakthrough in the field of AI Agent. An analysis article by Tencent Cloud Developer Community positioned it as "AI 2.0" and believed that it represents the evolutionary trend from "dialogue tool" to "long-term execution system". The article points out that competition among AI Agents in the future will focus on three aspects: who remembers more, who does things more stably, and who can optimize in the long term.   Compared with competing products, Hermes Agent and Claude Code have different positioning. Claude Code focuses on coding assistant functions and is suitable for short-term tasks; Hermes Agent is positioned as a long-term AI system and is suitable for scenarios that require continuous learning and accumulation. The two are not directly substitutes, but complementary.   The Chinese community’s attention to Hermes Agent is also increasing. The official Chinese community website (hermes-agent.org.cn) provides complete localized documentation and installation tutorials, including Windows WSL2 installation, long-term memory configuration, Skills, MCP, message gateway, multi-platform access and self-hosted deployment, etc. The community also provides WeChat groups and quick start guides, which lowers the threshold for Chinese developers.

• 7. Controversies and Risks (809 chars)

  In April 2026, the Chinese AI team EvoMap publicly accused Hermes Agent’s core self-evolution capability of being a systematic copy of its Evolver engine. This accusation triggered a big discussion in the open source community about the boundaries of technological innovation. Nous Research responded twice and denied the relevant accusations, saying that its technology is completely self-developed. The controversy reflects a growing focus on intellectual property and originality in the open source field of AI.   From a technical risk perspective, Hermes Agent is highly dependent on external APIs and network access, and users need to solve API stability issues by themselves. Some users reported that they encountered problems such as unstable API calls and limited network access during use. In addition, the skill sharing mechanism also has security risks, and users need to verify their source and security when introducing third-party skills.

• 8. Applicable groups and usage suggestions (929 chars)

  Hermes Agent is suitable for the following types of users. The first is the enterprise technology team that needs team knowledge management. AI can remember all technical decisions of the project, and new members can directly check the historical accumulation when they join. Secondly, developers who need to automate daily tasks, such as scheduled reports, data backup and other repetitive tasks can be automatically completed by Hermes Agent. The third is in-depth users who hope that AI can adapt to their personal work habits. After long-term use, AI will remember the user's coding style, preferences, etc.   Unsuitable scenarios include scenarios with extremely high data privacy requirements (although local deployment is supported, the configuration is highly complex), native Windows users (currently less stable than Linux and macOS), and users unfamiliar with self-hosted operation and maintenance.   As for alternatives, if you only need simple coding assistance, you can consider Claude Code; if you need a more lightweight solution, you can consider the Ollama local model solution.

• 9. Summarize (493 chars)

  Hermes Agent represents an important trend in the evolution of AI Agent from "conversation tool" to "self-evolving system". Its built-in persistent memory and automatic skill generation capabilities make AI no longer a "goldfish" that is cleared after each conversation, but a "digital partner" that can grow with users. For users who need to run AI tasks for a long time and accumulate team knowledge, Hermes Agent provides one of the most mature solutions currently. With the improvement of the Chinese community, the threshold for domestic developers is also constantly lowering.

No in-depth report yet

• 

  KDavis369213

  —

  After using it for a week, I feel it is more suitable for long-term projects than OpenClaw. The memory function is really good. I asked Agent to help me optimize Python scripts before. Next time I talk about similar topics, it will remember the previous demand background without having to explain it again.

---

• 

  AnthonyScott_77

  —

  Deployment is a bit difficult, so novices are not advised to rush in directly. It is recommended to read the tutorials in the Chinese community before starting.

---

• 

  Raymond.Lewis_Max

  —

  I regret not using it sooner! The scheduled task function is so convenient. It can be set to automatically generate data reports every morning, which saves a lot of time.

---

• 

  Charles.HarrisK

  —

  Finally supports Feishu! Previously, only Telegram and Discord were available, but now domestic platforms can also be used, I’m so touched!

---

• 

  周艳

  —

  Comparing it with Claude Code, the positioning is different. Claude Code is more suitable for short-term tasks such as writing code, while Hermes is suitable for scenarios that require long-term accumulation.

---

• 

  SatsChaserPowell

  —

  The memory usage is a bit large, and the fan spins wildly when running. It is recommended to add a memory optimization mode.

---

• 

  uwf4mobo17_q

  —

  The skill system is well received. The SKILL.md automatically generated by the Agent is of high quality and can be used after a few debuggings.

---

• 

  书生858

  —

  The Chinese support is indeed much better than other open source agents. It doesn’t have that machine-to-machine feeling and is comfortable.

---

• 

  Olivia97

  —

  The scheduled task function is really easy to use. You can set up a task to automatically back up the database every week, so you no longer need to do it manually.

---

• 

  JanetReedZ

  —

  yyds！

---

• 

  Lydia.Turner_Pro

  —

  After using it for a month, the memory function is really powerful. But I feel that the skill loading is a bit slow, especially when the project becomes larger.

---

• 

  N_odeLink

  —

  Does anyone know how to configure multi-model switching? I want it to use the cheap model for simple tasks and switch to Claude for complex tasks.

---

• 

  烟雨332

  —

  It is recommended that the official one-click deployment script for Docker is released. Manual configuration is too troublesome now.

---

• 

  Lawrence.Young520

  —

  After using it for a month, my true feelings are: I’m not bragging about it, and it’s really cost-effective. Chinese text processing is very natural, much better than those English AIs with translation accents.

---

• 

  NatalieClark

  —

  Help! Can anyone tell me how to turn off the automatic creation of skills? Some requirements do not want to be recorded as skills.

---

• 

  DomeniqueQuaak

  —

  The deployment has been running on the server for a week and is stable! He just needs to maintain the server by himself, so Xiaobai persuades him to quit.

---

• 

  Donald.TorresX

  —

  What is better than OpenClaw is the memory system. OpenClaw is a new session every time, and Hermes can remember what has been talked about before.

---

• 

  37VQ9

  —

  Compared with Notion AI, Notion is more focused on document processing, while Hermes is more suitable for development scenarios, and each has its own uses.

---

• 

  Shirley_Robinson007

  —

  Using the `/compress` command to compress the context can really save tokens, learned the lesson!

---

• 

  Walter.Mitchell

  —

  The memory function is really great. I asked the agent to optimize a Python script before, but when I looked for it next time, it actually remembered what I said last time about processing logs, and directly added error handling based on the previous one.

---

• 

  Ann.White_Max

  —

  Thanks to the Chinese community for the tutorial, which saved my life!

---

• 

  范贞月

  —

  Free + open source + self-hosted, these three combinations are what you need.

---

• 

  Carl346

  —

  The only disadvantage: the response speed is not as good as commercial products. After all, it has to be run on its own server and depends on the network and model response.

---

• 

  CarolFoster

  —

  It’s great to have access to corporate WeChat! Finally, you can summon AI assistants on WeChat at any time.

---

• 

  lazycat675

  —

  I finally understand the difference between OpenClaw and OpenClaw. OpenClaw is more like a conversational agent, while Hermes is a system-level agent that can perform scheduled tasks by itself.

---

• 

  George_Chavez_Pro0

  —

  Compared with OpenClaw, Hermes' learning ability is indeed unique. OpenClaw is strong in pure programming, but Hermes can remember what you have been taught.

---

• 

  SEped2025

  —

  After using it for two weeks, the skill accumulation is really useful. If a bug that has been solved before is encountered again, the solution can be reproduced by directly adjusting skills.

---

• 

  KellyPerryQ

  —

  Just deployed, it feels like it has too many functions for ordinary users. In fact... it is suitable for developers.

---

• 

  EJames_776

  —

  I would like to ask everyone, which Hermes model is the most cost-effective? GPT-4 is too expensive and I want to find a cheaper alternative.

---

• 

  夏浩

  —

  Memory function yyds! I asked it to help me process logs before. This time, error handling has been added to optimize performance, so there is no need to repeat the background at all.

No reviews yet