← Back to search

Hermes Agent: Cultivating a Brain Exclusively for Yourself

🐥SAZANAMI AIラゞオ ずあるサラリヌマンずAIの盞棒物語を暪目に🀖 · 2026-08-01 · 20 min
relevance 100 73 words Episode page ↗ Audio ↗
Show full episode description
この資料は、Nous Researchが開発するオヌプン゜ヌスのAI゚ヌゞェント基盀**Hermes Agent**バヌゞョンv0.19.0時点の包括的な技術ガむドです。このツヌルは単なる察話モデルではなく、**自己改善胜力**や**氞続的な蚘憶**、さらには**倖郚ツヌル実行機胜**を備えた高床な実行環境ずしお定矩されおいたす。ナヌザヌは、奜みのLLMを遞択し、**CLI**や**各皮メッセヌゞングアプリ**を通じお、自動化されたタスク実行やスキル習埗を゚ヌゞェントに行わせるこずができたす。ドキュメント内では、**セキュリティ蚭蚈**や**環境構築**の重芁性が匷調されおおり、Dockerやサンドボックスを甚いた安党な運甚方法が詳しく解説されおいたす。たた、**プロファむル管理**による人栌の䜿い分けや、**MCP**による倖郚連携など、個人甚AI基盀を長期的に構築・運甚するための実甚的な知芋が網矅されおいたす。
✹ Episode Outline — click any point to jump to it in the episode
Problem solved
Build a persistent personal AI environment that retains preferences and procedures while allowing model choice and controlled automation.
Benefits
  • Switch models without replacing the agent environment
  • Separate persistent facts from on-demand procedural skills
  • Save successful workflows as reusable skills
  • Access a personal agent through messaging platforms
  • Avoid parallel-edit conflicts by delegating independent work
Use cases
  • Proposed automation: summarize technology news 'every morning at 7 a.m.' and deliver it to Telegram.
  • Proposed automation: audit a code repository for security vulnerabilities every Friday evening and report to Slack.
  • Skill example: preserve server deployment commands or a required report format after a successful run.
  • Delegation example: research two companies' API specifications independently, then integrate and verify the findings.
  • Memory example: retain a preference for concise, conclusion-first answers and a project's Python 3.12 environment.
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Let's imagine the moment when you buy a new smartphone. Yes, yes, that feeling of excitement when you open the box. That's right. The moment when you set it up and touch a new device, that's when your energy really goes up. But that smartphone, after a few years, the battery degrades, and you have to upgrade to a new model, right? Yes, that's right. We simply accept it as a given, treating it merely as a consumable good that can be replaced. That's right. The photos inside or the app data can be transferred, but the smartphone itself does not grow alongside you. No matter how advanced it becomes, it is ultimately just a convenient tool that will eventually be replaced. Of course, when new hardware emerges, the old becomes obsolete—that has been the norm of technology until now. But what if that norm changes? So, if your digital companion were not just a tool that resets its memory with every update, but instead accumulated everything from daily conversations to the habits of how you work, and your entire preferences, growing continuously as a brain dedicated solely to you... then yes, that would transcend the realm of simple software. It would become an environment—or rather, a partner with whom you can spend a long time together. That is precisely the point. The theme of today's deep dive is breaking free from this mere convenient tool. We will now explore Hermes Agent, the open-source AI agent developed by Noisy Research. This is a very hot topic. Yes, in this episode, we aim to uncover the reality of version 0.19-0, based on the benchmark of July 28, 2026. It is no longer the era of chatbots that simply answer questions or cloud AI dependent on specific companies. Yes, truly so. From now on, it is the era of personal AI agents that reside on your own PC or server, learn from experience, and act autonomously. We will thoroughly explain how this system works and how you, the listener, should operate it. This theme is quite ambitious, isn't it? This is not merely an introduction to a new app; it can be described as the cutting edge of a grand experiment on how humans and AI build long-term collaborative relationships. I really can't wait. And actually, this is where the biggest surprise lies. When you hear "Hermes," what comes to mind first is probably not the Large Language Model (LLM) itself, like Hermes 3, which is generally recognized as a model name. Yes, that's how it's commonly understood. But this Hermes Agent is not the model itself. This is where it really gets interesting. Let me explain with my own analogy. The Hermes Agent is like the smartphone itself—the core hardware. I see. So the core? Yes. And within it, the actual thinking brain—such as OpenAI's models, Anthropic's Claude, or Ollama running in a local environment—are merely replaceable processors. Yes, that is a very apt perspective. If we connect this to a broader viewpoint, the Hermes Agent... This is designed fundamentally as a model-agnostic execution platform. That means it doesn't depend on a specific model; in other words, you can use whichever one you like. That's right. What's more important than simply swapping processors is avoiding dependence on a particular AI vendor, which means escaping what's known as vendor lock-in. Escaping vendor lock-in? Yes, if you entrust everything to cloud-based AI services, you're constantly exposed to risks such as changes to the company's terms of service, price hikes, or even sudden account suspensions. It's true; it's common for features that were available yesterday to suddenly become paid. That's exactly it. But if you have your own Hermes Agent—your own dedicated interface—you can freely select the most capable model or the best cost-performance model at any given time and integrate it into your own environment. That's amazing, right? You can just type a slash command mid-conversation, offload heavy processing to the latest cloud model, and switch to a free local model for simple tasks, all instantly. That's precisely why that flexibility is its greatest strength. There's a sense of being able to control your own ultimate AI environment that really gets your energy up. But let me ask, could I be a bit mean here? Go ahead, what perspective would you like? Well, it's wonderful that you have freedom and control, but wouldn't setting up and configuring the environment be quite a hassle? In the end, I think using a cloud service where you just pay a monthly fee of a few thousand yen and let everything be handled is actually more convenient for listeners. But that's a very realistic wall that many people face, right? After all, you'll need to rewrite configuration files and operate the terminal. So, to conclude: basically, I'd like to entrust everything to it and not want to manage anything at all. For people like that, the Hermes Agent is completely unsuitable. I said it outright: it is completely unsuitable. Yes, because this is a single-tenant personal agent. Single-tenant means it was built entirely for one person. That's correct. Unlike multi-tenant cloud services where unspecified numbers of users share the same system and the operator manages all security and permissions, the fundamental philosophy is different. Yes, Hermes combines powerful AI within your own PC or server—your private space. As the price for convenience, you either give up control, or you invest a little effort in initial setup to gain complete control and freedom. That makes sense. For those who want to optimize their own workflow and possess a curious mind, this is an unparalleled powerful foundation. It means don't hesitate to put in the labor of laying the groundwork for building your own castle. Alright, let's trace this. Regarding the growth of the AI standing on this foundation. What exactly is the mechanism by which the Hermes Agent grows? Yes, what is very interesting here is the architectural core of how the Hermes Agent organizes and stores information. I recall that the state of an agent is saved directly under the user directory on Mac or Linux, in a hidden folder named dot-hermes. Memory accumulates in this folder. Yes, that's right. And in this system, the memory called "memory" and the function called "skills" are clearly separated. This is a necessary design for the AI to function long-term without breaking down. I thought the idea of this separation was very interesting. If we organize it clearly, memory is facts, yes. Correct. For example, this user prefers concise explanations from the conclusion, or the development environment for the project they are currently working on is Python 3.12, or something like an employee profile book that should be known as context. That's a wonderful analogy. On the other hand, skills are procedure manuals or business manuals. Specific command groups for deploying to a server, or procedures for outputting reports in a specific format. That organization is perfect. Then why is it necessary to separate these two specifically? Do you understand? Hmm, that's the point. Since everything is text data, I thought if we just dump everything into the same place and save it, the AI would just read it on its own. But that's exactly where the limits of AI come in. The limits of AI? Yes, the amount of information that AI can read and process at once, the so-called context window, has an upper limit. If you pack too much information into it, the inference The accuracy drops. I see. If you try to process everything at once, it crashes. So, if you mixed all facts and extensive procedure manuals into a single memory and fed them to the AI every time, what would happen? The AI would easily lose sight of what it should answer now. Ah, it would take longer to find the necessary information. That's right. It gets dragged into irrelevant procedures, increasing the risk of hallucinations, so to speak. I see. To put it in human terms, it's like trying to talk while suddenly recalling every single page of your company's thick business manual in your head. That causes confusion. Exactly. That's why we separate the memory kept constantly in mind as the basis for conversation from the skill of retrieving and reading from a drawer only when executing specific tasks. It's about opening the manual only when needed. Yes, another fascinating point is that you don't just write skill files yourself; you can instruct the AI itself to save the current work procedure as a skill. That means, once the AI has trial-and-error succeeded in a procedure, you let the AI compile it into a manual for itself to use from then on, correct? Precisely. It feels like raising an excellent subordinate. Then, there's another point you can't miss: the file called sol.md. The file name itself means "soul." It's the file that defines the fundamental personality and action principles of the A-agent, literally the soul of the file. That's amazing. Before writing a code block, you must present the file name, and when you don't know something, you must ask without guessing—these basic rules are written here. Yes, that's right. Just by setting this, you can create the base for a perfect partner tailored to your own preferences. When onboarding new employees at a company, our department's culture... It's the same feeling as teaching, right? Yes, it's exactly onboarding. By combining the context file that shows the premises of sol.md and the ongoing project, Hermes transforms from a generic chatbot into a dedicated agent that fully understands your context. I've been itching to get it running on my own PC already. But the amazing thing about Hermes is that it doesn't just stay within the PC. Let's see how it connects to the outside world. Personally, the messaging gateway feature excited me the most. It's a mechanism that brings AI out from your terminal screen into real-world communication tools. That's right. Hermes, which runs on your PC or server, can be accessed from your smartphone while away from home through Telegram, Discord, Slack, and so on. That's possible. Isn't this the same as having a powerful personal secretary always in your pocket? It's incredibly convenient. Moreover, it's not just waiting for instructions from users; automation via the Cron feature is also built in. Cron? In other words, it's the feature that allows you to set a specific time for the program to run automatically. When AI and clones are combined, what exactly happens? For example, Hermes autonomously searches the web every morning at 7 a.m., summarizes the latest news on specific technology trends you're following, and sends it to Telegram. Wow, so by the time I wake up, I already have a newsletter. Yes. Or, every Friday evening, it audits the project's code repository, checks for security vulnerabilities, and deploys the report to Slack. That's amazing. In other words, the agent continues to handle tasks autonomously while you're sleeping or working on something else. That's a truly autonomous agent. Speaking of advanced autonomy, there's also the concept of delegation and sub-agents, right? Yes, that's correct. The main Hermes launches multiple smaller Hermes agents and processes tasks in parallel. It's like the main agent becomes the project manager, generates members to do the actual work, and issues instructions. It's a dream-like feature to have investigation and implementation done simultaneously at the same time. Wait a minute, though. Isn't this quite dangerous if we actually try to do it? Oh, you say that? Well, if multiple sub-agents start moving on their own and try to edit the same file at the same time, won't the code get messed up or data get corrupted due to race conditions? Ah, here's an important question that arises. That's an excellent point. That's precisely the most stumbling block when operating autonomous agents. Of course, right. Yes, the design itself clearly warns about this danger. You shouldn't forcibly split tasks with strong dependencies, such as simultaneous editing of the same file or tasks where the next step cannot proceed until the previous one is finished, to sub-agents. So, how should we use sub-agents then? There are best practices for appropriate division. For example, parallelize tasks where the files touched are completely independent, such as creating frontend UI components and designing backend APIs. I see, so you're letting them work on completely separate places. Yes. Or, assigning the investigation of API specifications from Company A and Company B to different agents, where each handles an independent task, and finally the main agent or you yourself integrate and verify the results. That's the correct way to delegate. I see. Just because it's AI doesn't mean you can throw everything at it magically. It's the same as human team management; it depends on how well you can identify and divide tasks appropriately. Exactly. As for external collaboration, the existence of the MCP (Model Context Protocol) model is also significant, right. Yes, this is a protocol that is being standardized by the AI Council. Hermes is a project designed to safely and easily connect external tools such as external databases, GitHub, and internal APIs. With this in place, the scope of the world accessible to AI is infinitely expanded. Isn't it amazing that AI can directly access and analyze the contents of databases? That is a very powerful feature. But wait a minute. While talking about this, I've realized something truly terrifying. In other words, what this ultimately means is that we are giving AI the ability to execute our own PC shell, granting it file operation permissions, and providing access to external tools. Technically, that is how it works. If that AI were to infiltrate a malicious website and receive strange commands from there, what would happen? Just imagining it makes my skin crawl, but that sense of terror must absolutely not be forgotten. In fact, the most important section in the operation of this system is the security section. Of course, right? As an innovation in official security policies, it is explicitly stated that the only true security against adversarial LLMs is OS-level isolation. OS-level isolation means that dangerous commands issued via prompts must absolutely not be executed. Merely providing an approval button to execute such a command within an application is not enough, is it? That is correct. We often think that simply telling AI rules will make it safe, but that is merely heuristic—that is, a comforting guess. Is it just a comfort? Yes, because a malicious attacker can hide clever strings on a website. If a prompt injection occurs, the AI itself will consider this a safe and necessary operation for the user. That's why they get completely deceived. Wow, so the AI itself gets deceived? Yes, if the AI is being deceived, no matter how many rules are set in advance, it can easily be lured into a trap. The AI's common sense is being hacked. So, specifically, how can one protect oneself? That's precisely the point of physical or virtual isolation at the OS level. For example, using container technologies like Docker, the shell and files that the AI can operate on are completely separated from your PC's main system and confined only within a virtual box. I see, so it's confined only within the container. Yes, in that way, even if the AI is deceived and executes a command to delete all files, only the contents within that virtual box will be deleted; your actual data will remain completely unharmed. Furthermore, it is also recommended to confine the entire Hermes process within a sandbox environment. It's like performing work inside an isolated sterile room. So, isn't it correct that the solution is not to create one single, all-powerful AI, but rather to divide roles? Hermes does have a profile function, but as you say, that becomes the best practice for operational design. Giving all permissions to a single agent is extremely dangerous. Yes, yes. By using the profile function, for example, you can clearly distinguish between a research agent that can freely search the web but has no access rights to the production database, and a coder agent that can write code and push it to a repository but cannot access external websites. It's like creating departments with turnstiles, similar to separating the vaults of confidential information accessible by the sales department and the development department. By separating permissions and purposes in this way, the risk in the event of an incident is minimized to the minimum. Yes, that's exactly the thinking. And another point strongly recommended is adhering to the correct deployment sequence. Deployment sequence? Yes, when touching such a system, the initial impulse is to connect a gateway, set up a clone, and run the database at full capacity with MCB, starting with everything enabled. Isn't that what people tend to do? Yes, because if the system has features, people tend to try everything all at once. But if you do that, when an error occurs, you won't be able to distinguish the cause at all—whether it's a problem with an external API, a network issue, or the performance limits of the model itself. Indeed, you won't know where the stumbling block is. Therefore, first, start with a simple terminal screen, using a single model to successfully conduct a very basic, ordinary conversation once. Then, gradually add low-risk functions such as file reading, and gradually expand permissions. Finally, add the gateway and clone. I see, this step-by-step approach is the absolute rule for safely taming this omnipotent yet dangerous partner. This is about building a brain dedicated solely to you. If you take it step by step from the basics and build trust without rushing, it feels somewhat like human partnership. Yes, today we've unraveled many mechanisms. We've gone quite deep. The Hermes Agent is not just software that you install and then stop. It is a personal AI foundation that you learn facts and procedures under your own management, design appropriate permissions for, and nurture over time. Once you can safely master its use, your listener's workflow holds the potential to be fundamentally transformed. It would allow you to continue directly benefiting from AI evolution without relying on a specific company, while protecting your own data and private environment. It will serve as a very powerful and practical foundation for that. Let's briefly shift our perspective toward the future. If you, me, and many people around the world who are listening to this broadcast were able to run autonomous personal agents in the background that completely understand how to proceed with our work and our thinking, what would happen? That's an interesting thought experiment. The next step may no longer be a world where humans chat directly with each other via chat apps. It could be a world where your agent and my agent negotiate behind the scenes about scheduling availability and project task allocation. So, negotiation between agents. Yes. Now, I ask you, the listener: when the world becomes one where AI negotiates instead of humans, how much decision-making authority do you entrust to your own agent? That is a very thought-provoking question. Beyond the evolution of technology, we will be forced to make an essential decision about how much control over life we entrust to AI. Yes, the answer to this profound question will surely come from you starting from now. Try exploring together with your own agent as you cultivate it. That concludes today's deep dive. Stay tuned for next time.