Five words you need to know
Provider, Tools, Skills, Memory, Learning Loop. Two sentences each. You'll hear them everywhere.
Hermes has its own vocabulary. Five words show up over and over, and once you know them the rest of the docs stop feeling foreign. We'll cover each in depth later — for now, just get the shapes in your head.
Provider
A provider is the language model Hermes talks to when it thinks.
Hermes doesn't ship with its own brain. You plug in a provider, and your provider gives Hermes its reasoning ability. Nous Portal is the default, made by the same team that builds Hermes. OpenRouter gives you access to hundreds of models with one account. You can also point Hermes at OpenAI, Anthropic, DeepSeek, a local model running on your GPU — anything that speaks a compatible API.
You can switch providers with hermes model. This is a one-minute operation, not a migration.
Tools
Tools are the things Hermes can do with its hands.
Reading a file is a tool. Running a shell command is a tool. Searching the web is a tool. Sending a Telegram message is a tool. Hermes ships with a few dozen of them, grouped into toolsets you can turn on and off: web, terminal, file, browser, vision, image_gen, and more.
When you start a conversation, Hermes tells the provider which tools are available. The provider decides when to use each one. You don't write code to invoke tools — the agent does it for you, based on what you're asking it to do.
Skills
Skills are reusable pieces of know-how you install into Hermes.
Where a tool is a low-level capability (run a command, read a file), a skill is a package of instructions that tells Hermes how to do something well. "How to open a GitHub PR following our team's conventions" is a skill. "How to fine-tune Llama 3 with axolotl" is a skill. "How to generate an architecture diagram in Excalidraw" is a skill.
You install skills from a hub (hermes skills install <name>) or write your own. Once installed, you invoke one with a slash command in chat. Skills are how Hermes learns your workflow.
Memory
Memory is what Hermes remembers between conversations.
Hermes keeps two small, curated notes under ~/.hermes/memories/: a MEMORY.md where it writes down things it learned about your environment and conventions, and a USER.md where it keeps a short profile of you — your preferences, your style, how you like to be talked to. Both are hard-capped in size on purpose. Memory is supposed to be lean, not a dumping ground.
On top of that, Hermes stores every past session in a local SQLite file so you can search your history later.
Learning Loop
The learning loop is the engine that ties the four things above together.
Hermes isn't static. After each task, it automatically reflects: was there anything worth remembering from this conversation? Is there a recurring pattern that should become a Skill? Does an existing Skill need updating based on feedback?
This loop runs on its own — you don't trigger it. The result: the more you use Hermes, the better it knows your habits, the sharper its Skills get, and the faster it works. The first time you ask it to write a script, it might feel generic. By the tenth time, it already knows your naming conventions and error-handling preferences — nobody taught it, it figured that out on its own.
You'll see this mechanism in action later in the Memory and Skills steps. For now, just know that Hermes is an agent that grows with use.
How they fit together
Here's the one-sentence version of the whole system:
Hermes uses tools and skills to do things, talks to a provider to think, keeps memory so it doesn't forget you, and gets better over time through its learning loop.
If that sentence makes sense, you're ready to install it. That's the next step.