QUICK ANSWER:
Hermes Agent is a free, MIT-licensed AI agent from Nous Research that keeps a persistent memory of its own work in a local database and writes new skills for itself as it goes. It is among the busiest agents in the world by token volume, and that is worth planning for: the memory rides along in every request, so fixed overhead can account for the bulk of each call. It suits repeating, long-running workflows, and it pays to set up caching before you scale up.
Almost every AI tool you use is exactly as useful on day one hundred as it was on day one. You explain your business again. You paste the same context again. You correct the same mistake again. Hermes Agent was built specifically to stop that, and it is the reason the project went from a February release to one of the most used pieces of open-source AI software in the world inside six months.
What makes Hermes different
Hermes Agent does not start from zero every session. It keeps what it learned about your business, your files and your preferences in a database on your machine, and it writes new skills for itself when it hits something it could not do.
Nous Research publishes Hermes Agent under an MIT licence. The architecture has three parts worth understanding before you install it.
The first is a layered memory system that persists between sessions. Conversation and state live in a local SQLite database in your home directory, which means it is yours, it is inspectable, and it is backed up by whatever backs up that machine.
The second is the learning loop. When it completes a task, it can distil what worked into a reusable skill, so the second time you ask for the same kind of job it has a procedure rather than an improvisation. This is the part that makes month three feel different from week one.
The third is reach. It connects to more than twenty messaging platforms and runs work across several execution backends, from your local shell to Docker, SSH and remote sandboxes. In practice that means you can start a job from a phone and have it execute somewhere sensible.
How fast it grew, and why that matters to your budget
Hermes went from launch to one of the most starred projects on GitHub in a matter of months, and it has held a place at the very top of OpenRouter’s daily token usage, at a scale measured in hundreds of billions of tokens a day across its user base.
That is usually quoted as a popularity stat. It is also a useful planning stat. It tells you these agents do a great deal of reading and writing per unit of useful work, so it is worth setting your provider up for that pattern before you scale usage.
The token overhead nobody puts on the landing page
This is the most useful thing in this article, so it gets its own section.
Analysis raised in the project’s own issue tracker found that the large majority of each API call was fixed overhead rather than your actual request, on the order of fourteen thousand tokens of system prompt, tool definitions and memory context riding along on every single call. The exact figure moves with your configuration and with each release, and the maintainers have been working it down. The shape of it is what matters: you are not billed for your question, you are billed for your question plus everything the agent needs to know in order to answer it.
Two consequences follow, and both are actionable.
First, cache pricing stops being a footnote and becomes the main event. Providers charge a fraction of the normal input rate for cached prefix tokens, and with Hermes the cacheable prefix is most of the request. Choosing a model with aggressive cache discounts, and configuring it so the cache actually hits, is the single biggest lever on your bill. Kimi, for instance, charges roughly a tenth as much for cached input as for uncached, which is a ten-to-one saving on the part of the request that never changes.
Second, batching pays handsomely. One job with twenty steps costs far less than twenty separate one-line questions, because the overhead is paid once rather than twenty times. Group your work and the economics improve immediately.
Getting it running
Setup is more approachable than the architecture suggests. The only hard prerequisite is git; the official one-line installer handles the Python and Node runtimes. A basic local install is a half-hour job. Putting it on a VPS so it can work while you are asleep is closer to an hour, plus the same hardening discipline any always-on agent needs.
You then point it at a model provider. It supports the Nous portal, OpenRouter, OpenAI-compatible endpoints and a long list of others, so you can run a cheap model for routine steps and a stronger one for the hard ones. There is also a migration path from OpenClaw, which is worth knowing if you already have skills and channels configured there.
Two workflows where it earns its keep
The recurring report nobody wants to write
A weekly performance summary that pulls from the same four places, formats the same way and lands in the same inbox is exactly the kind of job that gets better on repetition. The first run needs correction. By the fourth, the agent has the procedure stored and produces something you edit rather than rebuild. This is where a learning agent beats a chat window decisively, because the corrections stick. If that report is a marketing one, it pairs naturally with the reporting we set up as part of search campaign management.
The long research job
Work that runs for hours, keeps notes and picks up where it left off suits a persistent agent far better than a session-based assistant. Competitor monitoring, tender scanning and market research all fit. Just be deliberate about scope, because a research agent left running is also a token meter left running.
Hermes Agent compared
| Tool | Licence | Memory | Best for |
|---|---|---|---|
| Hermes Agent | MIT, free | Persistent, local, self-improving | Repeating workflows that should compound |
| OpenClaw | Open source, free | Session and skills based | Broad action taking across many channels |
| Managed assistants | Per seat, per month | Vendor-held, limited | Teams that want zero maintenance |
The honest comparison with OpenClaw is that they overlap heavily and differ in emphasis. OpenClaw is about reach and action. Hermes is about accumulation. If you do the same fiddly thing every week, Hermes is the better bet. If you mainly want to trigger things from your phone, either will serve you well.
Pros and Cons
Pros
- MIT licensed and free, with no per-seat cost.
- Memory persists locally in a database you can inspect and back up.
- Gets measurably better at repeated tasks rather than staying flat.
- Works with more than twenty messaging platforms and many model providers.
- One-line installer, so a working local setup is a half-hour job.
Worth knowing before you start
- Fixed token overhead per call, so batching beats lots of small requests.
- Worth running a metered trial month before you forecast the budget.
- Fast-moving release train, so pin versions and read the notes.
- The stored memory deserves the same care as a customer database.
- A young and fast-moving project, so expect frequent improvements.
The verdict
Hermes Agent is the right choice if you have workflows that repeat and you are prepared to invest the first fortnight teaching it. The memory system is not marketing; it changes what the tool is good for. Go in with the cost model in your head: batch your requests, use a provider with real cache discounts, and set a spending cap on the API key before you start. Do those three things on day one and it is one of the best value tools in this category.
Work out whether an agent actually fits your business
An agent that learns your processes is only worth building if your processes are worth repeating. That is the assessment worth doing first, and it takes a conversation rather than an install. Our AI consulting service starts with an audit of where your team’s hours actually go, and if you would rather build the capability in-house we run AI training workshops that cover exactly this kind of deployment.
It is free and MIT licensed, so a trial costs you nothing but the model calls. You can find it at Hermes Agent.
This guide describes how the pricing is structured rather than quoting figures that date: tiers, ratios and what each level unlocks all hold when rates change. For the current numbers, check the vendor’s own pricing page.
Frequently asked questions
What is Hermes Agent?
Hermes Agent is a free, MIT-licensed AI agent from Nous Research. It runs on your own hardware, keeps a persistent memory in a local database, and can write new reusable skills for itself based on work it has already completed.
Is Hermes Agent free?
The software is free and MIT licensed. You still pay for the model calls it makes and for any hosting. Because the agent carries a large fixed context on every request, the model bill is usually the dominant cost.
How does the Hermes Agent memory work?
Conversation and state are stored in a local SQLite database in your home directory, layered so that recent detail and longer-term learning are handled differently. Because it is a local file, you can inspect it, back it up and delete it, which also means it needs protecting like any other sensitive data store.
How do I keep Hermes Agent running costs down?
Every call carries the system prompt, tool definitions and memory context alongside your actual request, so a large share of each request is fixed overhead. Batching several tasks into one job and choosing a provider with strong cached-input discounts are the two changes that cut the bill most.
Should I choose Hermes Agent or OpenClaw?
Choose Hermes if your value comes from the same tasks repeating and improving over time. Choose OpenClaw if your value comes from breadth of action across many channels and tools. They are close enough that most people run one, not both.
