Hermes Agent: The Free AI That Never Forgets Your Business

Every AI chat you have ever had with your business has been thrown away the moment you closed the tab.

Ryan runs a six-person plumbing and heating firm outside Manchester. Every quote he sends follows the same private logic: measure the job, price the fixtures, add labour by the hour, add a call-out fee if it is more than ten miles out. He has never written that logic down. It lives in his head, and every apprentice who joins has to learn it by watching him do it for three months.

Last winter he tried using ChatGPT to speed up his quotes. It worked, for about twenty minutes. Then the session ended, and the next morning he was back to explaining, from zero, how he prices a boiler swap versus a full central heating install. He was not managing an assistant. He was re-hiring a stranger every single day.

You keep re-hiring the same amnesiac assistant

This is the part nobody warns small business owners about when they get excited about AI. A chatbot is not lazy. It is amnesiac by design. Close the tab, and the context, the corrections, the three months of "no, do it like this instead," all of it disappears with it.

So you end up paying, in time if not in money, for the same onboarding conversation over and over. Explain the pricing. Explain the tone. Explain that a Tuesday call-out costs more than a Thursday one. Every session starts from nothing, which means every session is worth less than the last one should have been.

An agent that actually remembers changes that arithmetic completely. That is exactly the bet a lab called Nous Research just shipped, and this week it is the thing every AI creator on X and YouTube cannot stop talking about.

What is Hermes Agent?

Hermes Agent is an open source, MIT licensed AI agent built by Nous Research, the lab better known for its Hermes family of language models. It first launched on February 25, 2026, and has shipped a fast run of updates since, reaching version 0.18.2 on July 7, 2026. Its tagline, straight from Nous, is blunt about what it is trying to be: "the agent that grows with you."

Unlike a chatbot, Hermes is built to run as a long-lived process, not a one-off conversation. It carries a persistent memory system built on SQLite with full-text search, so it can recall what you told it last week, not just what you typed a minute ago. And it is model-agnostic: it can drive more than 300 different models through OpenAI, Anthropic, OpenRouter, or a local install, so swapping the model underneath it does not wipe out anything it has learned about your business.

Why a memory beats a bigger model

Here is the part that matters more than any spec sheet. Practitioners have started calling the software wrapped around a model the harness, and a growing body of testing suggests the harness can matter as much as the model itself. One widely cited experiment kept the underlying model completely fixed and improved only the harness around it, the system prompt, the tools, the feedback loops, and moved its score on a hard benchmark from 52.8 to 66.5 percent. Same brain. Better scaffolding. A 13.7 point swing.

Hermes is that idea shipped as a finished product instead of a research paper. Its core mechanic is what Nous calls a closed learning loop, and it has three moving parts. First, skill creation: when Hermes works out how to do something, it writes the solution to a reusable skill file. Second, skill improvement: that file is not frozen, the agent revises it as it learns what works. Third, curated memory: Hermes periodically nudges itself to write down what is worth keeping, rather than letting everything vanish when the chat window closes.

For a plumber, a bakery, a bookkeeping practice, or a marketing agency, that loop is the entire pitch. You are not renting a smarter brain for twenty minutes. You are slowly building an asset that gets more valuable every time someone uses it.

Isn't this just a chatbot with a fancier name?

No, and the distinction is concrete, not marketing language. A standard chatbot session is stateless. Ask it your pricing logic today, and tomorrow's chat starts blank. Hermes ships with over 40 built-in tools, terminal access, file operations, browser automation, image generation, and natural language task scheduling, and it connects to more than 20 messaging gateways, including WhatsApp, Telegram, Slack, Discord, and Microsoft Teams. You are not opening a new tab. You are messaging the same coworker you messaged last month, and that coworker still remembers what you told it.

In June 2026, Nous Research shipped a feature called the /learn command that makes this even more direct for a small business. Point Hermes at a folder of source material, your SOPs, your price list, your product manuals, your old invoices, and it distills that material into a verifiable, reusable skill on its own. You are not writing a prompt. You are handing over the same binder you would hand a new hire, and the agent reads it once.

How do you actually set Hermes Agent up for your business?

You do not need to be a developer to get a basic version running, though a little comfort with software helps. Here is the shape of it, based on Nous Research's own release notes and a live setup walkthrough from AI creator Callum Wanderloots on YouTube.

1. Choose where it runs. Hermes is free to self-host on a Linux box, a Mac, or even a cheap VPS, since it is MIT licensed. If you would rather not manage a server, hosted options exist that run it in the cloud for a flat monthly credit fee, starting from roughly 20 US dollars a month for always-on use.
2. Introduce yourself. The first thing to do is tell the agent who you are and how you like things done. It writes that straight to memory, and every future session pulls it back automatically, no re-explaining required.
3. Feed it your business, not your prompts. Use the /learn command on your actual SOPs, price sheets, and past quotes. Hermes turns that material into a named, reusable skill file rather than a one-time answer.
4. Connect a messaging channel. Link it to WhatsApp or Telegram in a few minutes, so you and your team can talk to it from a phone instead of a browser tab.
5. Automate the recurring part. Ask it to run a task on a schedule, a daily summary of new enquiries, a weekly stock check, a morning briefing, and it sets up the automation itself, no cron syntax required.
6. Correct it, out loud. When it gets something wrong, say so in plain language. Hermes edits the underlying skill file itself, so the fix compounds instead of evaporating.

Chatbot versus a self-improving agent

What this looks like in Ryan's shop

Ryan does not need to rebuild his business to try this. He needs an afternoon. He points Hermes at a folder of his last fifty quotes and his supplier price list, and runs /learn. The agent comes back with a skill it calls quote-a-job, built from the actual pattern in his own numbers, not a generic template.

He connects it to the WhatsApp number his customers already message. Now, when a message comes in describing a job, Hermes drafts a quote using that skill and sends it to Ryan for a thumbs up before it goes out. The first week, he corrects it constantly, on call-out radius, on which fixtures need a premium markup, on how he phrases a decline. Every correction gets written into the skill, not just remembered for that one chat.

By week three, most quotes go out with a light edit instead of a full rewrite. Ryan has not hired anyone. He has, without quite meaning to, written down twenty years of pricing judgment in a form his newest apprentice, or his next AI agent, can actually read.

What you have after this

A week in, you have a skill file that reflects how you actually work, not a generic AI answer. A month in, that file has absorbed dozens of small corrections and needs far less babysitting. Six months in, you have something closer to a documented playbook of your business than most owners ever get around to writing themselves, sitting inside a tool that keeps using it every single day.

Same business. Same you. A memory that finally does not reset every morning.