Getting Started with Hermes Agent: Memory, Skills and What It Actually Costs

Paul — Sun, 08 Nov 2026 14:54:39 +0000

An AI assistant is more useful when you don't have to explain the same project every time you open a chat.

That's the appeal of Hermes Agent, an open-source agent from Nous Research. It can keep persistent notes, turn working procedures into reusable skills, and run tasks through tools. You can use it in a terminal or connect it to messaging apps such as Telegram.

There's a bit of setup involved, and "self-improving" needs some explanation. This guide covers the practical bits without pretending an agent can reliably handle everything unattended.

Disclosure: There's a Nous Portal referral link in the costs section. Eligible new subscribers get $15 off their first paid month, and I receive $10 in Portal credit for a qualifying referral.

What is Hermes Agent?

Hermes Agent is software that connects a language model to tools, saved context and an environment where it can work.

Depending on your configuration, it can browse the web, work with files or run terminal commands. Its GitHub repository documents the available interfaces and setup commands.

Hermes Agent isn't a single model. You choose a provider and model, and can change that choice later. Nous Portal is one option; the project also supports other providers and compatible endpoints.

That matters when comparing costs. The agent software is free and open source under the MIT licence. Model usage and optional hosting are separate.

Persistent memory, with limits

Hermes has built-in memory that survives between sessions. The memory documentation describes two main files:

MEMORY.md holds environment facts, conventions and things learned.
USER.md holds information about your preferences and expectations.

These are compact, curated notes rather than an unlimited transcript pasted into every conversation. Hermes also has session search for finding things in earlier chats.

A useful first request might be:

Remember that my project documentation lives in the docs folder, and ask before changing existing files.

Then check that the memory was actually saved. The docs warn that an agent can say it remembered something without calling the memory tool.

On messaging platforms, /new is handy at natural task boundaries. Restarting the gateway doesn't necessarily start a fresh conversation, and newly saved memory is loaded at the next session boundary.

Don't treat memory as a password manager. Keep credentials in proper configuration or secret storage.

What "self-improving skills" means

A skill is a saved procedure the agent can use again.

Say you ask Hermes to prepare a weekly research brief. You work out which sources matter, how links should look and what counts as relevant. Those instructions can become a reusable skill instead of something you explain again next week.

The project describes a learning loop that can create skills after complex tasks and improve them with later use. That's procedural learning, not the model retraining itself or becoming right about everything.

A procedure can keep a mistake just as easily as a good idea. Review important skills, especially ones involving shell commands or outside accounts.

For a first skill, pick something easy to check, like formatting notes. Don't start with unattended changes to production infrastructure.

Getting started locally

The official site offers desktop downloads for macOS and Windows, plus terminal install instructions.

For Linux, macOS or WSL2, the documented installer is:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

That command downloads and runs a script. Read the instructions and look at the script first if you're not comfortable running remote installers.

After installing, open a fresh terminal if needed and run:

hermes setup
hermes

The setup wizard configures your provider and other options. You can change models later with:

hermes model

For the Nous Portal route, the repository documents:

hermes setup --portal

This uses a browser-based login to connect Portal and set up its tool gateway.

Start with a small task in a throwaway folder:

Read these sample notes and draft a summary into a new file. Don't modify the originals.

Check the output and any tool activity. It's a better introduction than handing a new agent your whole home directory.

Talking to it through Telegram

Hermes has a messaging gateway, so you don't have to sit at a terminal to use it.

The repository lists Telegram, Discord, Slack, WhatsApp and Signal among supported platforms. Setup differs by service, so follow the messaging guide for yours.

The documented starting commands are:

hermes gateway setup
hermes gateway start

Telegram setup means configuring a bot and restricting who can talk to it. Take that seriously: a bot connected to terminal tools shouldn't be open to strangers.

The gateway also has to keep running. Messaging your agent from your phone doesn't move the work onto the phone. If Hermes runs on your laptop, the laptop still needs to be awake and online.

Cloud hosting helps when you want an always-on agent without leaving your own machine running.

Scheduling useful work

Hermes includes a scheduler for one-off and recurring tasks, and you can describe them in plain language instead of writing cron expressions.

For example:

Every weekday at 9am New York time, check these two documentation pages for updates and send me a summary with links. Don't publish anything.

A good scheduled prompt names its sources, defines the output and says where it should go. Include a timezone rather than hoping the agent guesses.

According to the scheduling docs, scheduled agent tasks run through the gateway in fresh sessions, so the prompt needs to stand on its own.

Check things with:

hermes cron list
hermes cron status

Test a task before relying on it, and check delivery too. A successful research run isn't much use if the summary never reaches you.

What does it cost?

Hermes Agent itself is free and open source. You don't need a Nous subscription just to install it.

Your bill depends on the model provider, tools and hosting you pick. Running locally avoids a hosting bill, but paid model calls still cost money.

Nous Portal bundles several parts together. At the time of writing, Plus costs US$20 a month and includes:

Hermes Agent usage within limits that reset daily, weekly and monthly.
One cloud agent with up to 4 GB RAM and 25 GB disk.
Access to the models in the Portal catalogue.

You can use any model through Portal, but pricier models eat into included usage faster. It isn't unlimited access to expensive models.

One more billing detail worth knowing: Portal API keys always use pay-as-you-go credits, not your included usage. A separate app using your API key isn't covered by Plus.

Bigger cloud instances and extra agents use credits too. Check the current plan details before committing.

The official site has the plan comparison, and Portal's info page shows current model rates.

Referral offer, clearly disclosed

If Plus suits you, here's my referral link.

Eligible new subscribers get $15 off their first paid month. On the US$20 Plus plan, that makes the first month about US$5, before any taxes or extra usage. After that it's the listed US$20 a month.

I get $10 in Portal credit for a qualifying referral, so weigh my recommendation with that in mind. Check eligibility and the final price at checkout.

A sensible first project

Pick one small job you already understand well.

Ask Hermes to summarise a folder of non-sensitive notes. Check its output, then save your formatting preferences. Once that works, think about making it a skill or scheduling it.

Keep publishing, purchases and destructive changes behind explicit approval while you learn how your setup behaves.

The useful question isn't whether Hermes can do everything. It's whether it can handle one repeatable job well enough that you trust the next run.

DEV Community: Paul