Hermes can now learn a new skill from anything you show it — a doc, a folder, or a job you just did — and keep it forever.
"The only agent with a built-in learning loop — it creates skills from experience, improves them during use, and builds a deepening model of who you are across sessions."
— Nous Research, the Hermes Agent README, June 2026
Before — every session, I'd re-explain how I do things.
How I publish a guide. How I pull my search data. How I deploy.
The agent would do it, then forget the whole recipe the moment we were done.
Next time? Start from scratch. Explain it all over again.
I was the only one who remembered how my own workflows ran.
Then Hermes learned to learn.
After — now I show it once.
I point it at the workflow and say one word: learn.
It writes itself a skill — a clean, reusable recipe — and keeps it.
Next time I just say "do the thing" and it already knows how.
Teach it once. It knows it forever.
Here's what's happening for the founders already running Hermes inside the Boardroom — agency owners, ecom founders, creators, solo operators.
You've seen the proof. Real people. Real results.
The next few minutes show you exactly how this works — and I tested it live, so you'll see the real skill it wrote.
So here's the deal.
Promise yourself one thing right now — you finish this guide, and you teach your agent one skill before you sleep tonight. Just one. The boring thing you do over and over. Because the moment your agent learns its first skill, it stops forgetting — and it only gets smarter from there.
The people sitting still are still re-explaining themselves to a tool. The people teaching their agent today are building something that compounds.
Be one of those people.
Teach it once today. It knows it forever. That's the whole shift.
One command turns anything you can describe into a skill your agent keeps. Here's what makes it more than a party trick.
A docs page, a folder of code, a job you just walked it through, or notes you paste. If you can point at it, Hermes can learn it. No special format.
It writes itself a clean recipe and saves it. Next session it already knows how. You never re-explain the same workflow again.
No coding, no template to fill in. Hermes authors the skill itself — proper steps, the gotchas, what not to do — the way a careful person would write it down.
Skills sit quiet until the moment they're useful, then load on demand. You can give your agent fifty skills without slowing down a single chat.
The same skill works from the dashboard, the command line, voice (Jarvis), and your messaging apps. Learn it once — use it from anywhere.
Thinking it? "Great, another thing to learn and maintain."
There's nothing to maintain. You don't write skills — the agent writes them. Your only job is to say "learn" when you notice yourself explaining the same thing twice. It's less work than what you do now, not more.
Here's the real shift. The old way, your agent was a goldfish — brilliant for one chat, blank the next. The new way, it builds muscle memory.
Thinking it? "I already use MCP servers and plugins — isn't this the same?"
Those bolt on tools the agent can call. /learn is different — it captures how YOU do something into a recipe the agent keeps. Your workflow, your steps, your gotchas. You can't download that from anywhere — only your agent can learn it from watching you.
It's one command. You give it a source — Hermes does the rest, using the tools it already has. No separate "learning engine," no extra setup.
Here's the real thing. I pointed Hermes at a simple workflow and said /learn. Twenty-five seconds later it had written a clean, reusable skill — on its own:
And this is the actual skill it wrote — I didn't touch a line of it:
Thinking it? "Sounds like marketing. Did this actually run?"
It did — on my own machine, on a local free model, in 25 seconds. The skill above is the real file it wrote, copied straight out. No edits. That's the whole point of showing it instead of just telling you.
This was the first thing I checked. The same /learn runs from every surface — the dashboard, the command line, your messaging apps, and the Jarvis voice assistant. They all hand the job to the same agent. So you can literally talk to Jarvis, walk it through something, and say "learn that."
Thinking it? "My agent runs a small free model — will it actually pull this off?"
Be straight with yourself here. On a simple, self-contained workflow, my free local model wrote the skill in 25 seconds. On a harder source — "go dig through my whole repo and figure it out" — a small model can get lost and run out of steps. The fix is easy: point /learn at a clear source, or route the job to a stronger brain. The Agent OS lets you pick the model per job — so this is a setting, not a wall.
I run Hermes inside the Agent OS — one dashboard where my agents share one memory, one team, one screen. /learn is how I keep teaching it new tricks without writing a line of code.
link in the description ↗
Run hermes update so you've got the /learn command. (Mine was a few hundred commits behind — the update pulled it in.)
From the dashboard, the command line, or just talk to Jarvis. Any surface works.
Type /learn and then the thing — a docs URL, a folder, or "how I just did that." Plain words are fine.
Hermes reads the source and authors a clean SKILL.md on its own. Approve the write and it's saved.
Next time, just ask for the task by name. The skill loads on demand and your agent already knows how.
Thinking it? "What if it writes a bad skill?"
You see the file before it saves, and you can edit or delete it like any note — it's just a markdown recipe. If a skill's off, you fix one line or run /learn again with a clearer source. Nothing's locked in, and a wrong skill can't break anything else.
Wrong: "Teaching an AI a skill means writing code or config."
Right: You say one word — learn — and describe the thing in plain English. Hermes writes the skill itself. I watched a free local model do it in 25 seconds, no code from me.
Wrong: "My agent forgets everything, so why bother."
Right: That's exactly what this fixes. A learned skill is kept — it stops being a goldfish and starts building muscle memory. Teach it once, it knows it forever.
Wrong: "I'll wait until the tools settle down."
Right: An agent that learns gets further ahead every week you use it. Waiting just means a bigger gap between you and the people teaching their agent today. This compounds — start now.
3,600+ founders are building agents like this inside the Boardroom right now. Their wins — real businesses, real results — are documented here.
Read the 158-page member wins →You've seen it run — one word, a real skill, kept forever. If you want Hermes set up to learn your workflows, step by step, it's all inside the AI Profit Boardroom.
I'll see you inside ↗