Hermes Agent reference guide: installation, memory and skills, models, costs, and limitations

Hermes Agent is an agent from Nous Research that combines a language model, tools, persistent memory, and reusable skills to complete tasks. You can install it on your own machine, choose a model provider, and control its access. Its actual operating cost then depends on inference, the services it uses, and hosting.

What Hermes Agent actually does

The software receives a request, uses the available tools, examines their results, and continues working. One of its distinguishing features is its ability to retain facts and procedures between sessions. That continuity can save you from repeating instructions, but it also means checking what the agent has saved.

The official Nous Research repository publishes the code under the MIT license. The site’s October 2026 review records version v0.22.5, dated October 23, 2026. This identifies a specific state of the project; the commands in this guide follow the official documentation consulted during its preparation.

Keep the agent separate from the model it uses. Installing Hermes does not automatically install a local model. You configure the program that coordinates actions separately from the service that generates responses.

Hermes is intended for people comfortable managing a terminal, credentials, and configuration files. To see how this approach compares with other options, browse the AI agent directory. The article Hermes Agent: installation, costs, and limitations covers related ground; this page focuses on the decisions needed to keep an installation useful over time.

Prerequisites: prepare the environment and access

Before installing anything, decide where the agent will work and which data it can reach. For an initial trial, use a regular user account and a folder containing documents that are not sensitive. Avoid starting with your backups, production credentials, or complete work email account.

Prepare the following:

• an operating system covered by the official documentation;
• a terminal suited to your environment;
• network access to download the software and contact a remote API;
• a provider account or an accessible model server;
• a way to track spending with the provider;
• copies of the files you will use for testing.

Hardware requirements depend mainly on where the model runs. A remote API handles inference elsewhere; a local model requires hardware capable of holding its weights and supporting its context. The Hermes name alone does not imply a universal hardware specification.

Also define an initial objective with a result you can check: summarize a public document, find information in a test folder, or draft a procedure without executing it. A limited request makes troubleshooting easier. You can distinguish an installation problem from a model error or an unavailable tool.

1. Install using the official documentation

The commands below come from the official installation guide, consulted for this page on October 24, 2026. They download and execute a script. Check the domain and inspect the script if your IT policy requires it.

On Linux, macOS, or WSL2:

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

On native Windows, open PowerShell:

iex (irm https://hermes-agent.nousresearch.com/install.ps1)

These are alternative installation methods; use the one for your environment. The documentation also provides desktop application packages for macOS and Windows. Check hardware restrictions before selecting a package: the documented macOS package targets Apple Silicon.

Installing from source prepares the program and its dependencies. It may also download components for browsing and computer use. A completed installation does not establish that every optional feature is ready. Read the final installer messages and check for any failed downloads.

For a POSIX installation, the documentation places the code in ~/.hermes/hermes-agent/ and data in ~/.hermes/ by default. The Windows source installation uses a location under %LOCALAPPDATA%\hermes\. Keep that distinction in mind when backing up files or investigating a problem.

2. Configure the model and test a conversation

The official quickstart provides a setup wizard:

hermes setup

To select or change the provider and model interactively:

hermes model

The integrated setup option offered through Nous Portal uses this command:

hermes setup --portal

It configures access to the provider and its gateway tools. Review the billing terms before choosing this option. If you want to manage your own credentials, follow the corresponding setup process rather than assuming an existing subscription covers every call.

Then launch the interface:

hermes

Ask for a simple response, followed by a file read in your test folder. Check the displayed provider, whether the response makes sense, and whether the agent actually used the tool. Leave automation disconnected at this stage.

You can test resuming the latest session:

hermes --continue

The documentation consulted for this guide requires a minimum context of 64,000 tokens. With a local server, check the capacity actually configured, rather than relying on the model’s advertised capacity.

The Blank Slate configuration mode provides a reduced starting point, but still leaves file operations and terminal access available. It also disables memory and skills. Explicitly enable the features you need before testing whether they persist between sessions.

3. Understand and back up the configuration

The official configuration reference distinguishes ordinary settings, secrets, and authentication credentials. For the default POSIX installation, the main locations are:

Item	Location	Precaution
Settings	~/.hermes/config.yaml	Keep a copy before making changes
Keys and secrets	~/.hermes/.env	Exclude from public files and shared repositories
Authentication	~/.hermes/auth.json	Protect as you would access to the account
Memory	~/.hermes/memories/	Review the information retained
Skills	~/.hermes/skills/	Track procedures and their changes

The documented commands let you inspect settings and check for missing options or selections that have become outdated:

hermes config
hermes config check

To open the configuration file in your editor:

hermes config edit

Avoid changing several configuration layers at once. Change a setting, rerun a familiar test, and record the result. A typo in a key can leave a file readable while preventing the behavior you intended.

Keep backups separate from material you share. An archive that can restore your installation may contain keys, history, and personal information. When asking for help, prepare a small, redacted excerpt instead. Do not publish the entire data folder in a support issue.

Memory, history, and skills serve different purposes

Here, agent memory means selected information retained across conversations. It does not replace the full conversation history or your working files.

The official memory documentation describes two files: MEMORY.md for lasting notes and USER.md for user preferences and expectations. Their contents are loaded as a snapshot at the start of a session. A change saved during that session therefore does not immediately update the snapshot already in use.

Mechanism	Contents	Useful check
Current context	Exchanges and results from the active task	Does the agent have the right document?
Session history	Past conversations available for reference	Does it retrieve the right source?
Persistent memory	Selected facts and preferences	Is the saved fact accurate?
Skills	Reusable procedures	Does the procedure still work?

Memory has a capacity limit and does not automatically compact itself when a write exceeds that limit. The agent must then shorten, combine, or remove entries.

To test persistence, have the agent save a harmless convention, finish the task, and open a new session. For extended conversations through messaging services, the documentation recommends creating an explicit session boundary:

/new

Restarting the machine does not necessarily create that boundary. Also avoid having multiple processes write to the same data directory; use separate profiles. The article on AI agent memory examines the differences between retained information, retrieval, and context in more detail.

4. Check skills before reusing them

A skill contains instructions the agent can use for a task. The official skill system uses files including SKILL.md, with detailed content loaded when needed. Hermes can create, edit, and delete skills.

You can request an inventory of available skills with this documented command:

hermes chat --toolsets skills -q "What skills do you have?"

A procedure’s availability does not establish that it is appropriate. Before adopting it, examine its purpose, prerequisites, external actions, and verification method. A skill that publishes a document requires a different level of trust from a procedure that only formats it.

Autonomous skill creation can preserve a useful method, but it can also save a workaround specific to one situation or a command that later becomes obsolete. Here, “learning” does not mean the model’s parameters have been retrained. You need to be able to find the saved procedure and understand its effects.

For an internal skill, specify when to use it, the expected inputs, permitted actions, and stopping conditions. Test it on a copy of the data, using a normal case and an incomplete case. Check that it reports missing information instead of inventing it.

Finally, keep a copy of approved procedures so you can compare changes. The guide to AI agent skills explains this approach to writing and reviewing them.

Choose a model for the task and the data

The model configuration documentation explains provider and model selection. Base that choice on the quality of the agent’s actions, tool compatibility, context length, and the terms governing data handling.

OpenRouter is among the providers listed in the quickstart. A compatible endpoint can also connect to local infrastructure, including through Ollama. In every case, verify the address, exact model identifier, and capabilities exposed by the server.

Criterion	Remote API	Local inference
Operation	Service managed by the provider	Server and model you maintain
Expenses	Usage charges or provider plan	Hardware, electricity, and administration
Data	Content sent to the chosen service	Data flows to check within your infrastructure
Capacity	Depends on the available plan	Depends on hardware and configuration

A model that writes well may handle tools poorly. Assess whether it chooses an appropriate action, interprets failures, and stops when it lacks authorization.

Use the same small set of tasks to compare options: reading a document, extracting information you can verify, handling a deliberate access error, and responding to an ambiguous request. Record usable results and the interventions required. Changing models may also affect how Hermes selects memory and follows skills, so repeat those checks after switching.

Costs: budget for the complete task

The software license does not determine its operating cost. Your budget includes model access, external tools, infrastructure, and the time spent reviewing or repairing results. A monthly estimate that leaves out these items is incomplete.

For APIs billed by token, the basic calculation adds input tokens multiplied by their rate to output tokens multiplied by theirs. Any caching mechanisms and other contractual billing terms need separate treatment. The LLM pricing calculator helps you work through these assumptions.

An agent may make multiple calls for a single request. It adds observations, rereads context, attempts a correction, or pursues an unsuccessful line of investigation. A short final answer can therefore conceal substantial intermediate work.

To build a useful budget:

• record spending on representative tasks;
• count tasks accepted after verification;
• track failures and retries separately;
• add hosting and supporting services;
• compare spending with work actually completed.

Distinguish a spending alert from a limit that blocks further usage. Check with your provider what happens when the budget is reached, then test that behavior with a dedicated set of credentials. Also constrain open-ended research: a defined document list and expected result reduce unnecessary exploration.

Limitations and security: control the effects of actions

The official security documentation describes several protections, including user permissions, approval for dangerous commands, and terminal isolation. Their presence does not guarantee that an action is appropriate for your intended task.

A technically routine command can send the wrong file to the wrong recipient. Define approval requirements according to the consequences: publishing, deleting, paying, or changing shared data.

The documented configuration lets you choose a containerized terminal:

hermes config set terminal.backend docker

This command requires a working Docker environment. Check mounted directories, available secrets, and network connections. A container with broad access can still affect the surrounding environment.

Prompt injection presents another risk: a document the agent reads may contain instructions designed to redirect it. Treat pages, messages, and tool results as data, and scrutinize requests for additional access or data transfers.

Local memory may also be included in requests sent to the provider. Local storage and local inference are separate decisions. The guide to securing an AI agent expands on these checks. Start by keeping secrets out of working documents and granting only the access required for the task.

Common problems

The launch command cannot be found

Open a new terminal and check the installer messages. For Bash, the quickstart suggests reloading the shell configuration:

source ~/.bashrc

Do not repeat the installation with elevated privileges before diagnosing the problem. First check which account you are using and where the launcher was installed.

The program starts, but the model does not respond

Return to the interactive provider selection. Check authentication, available credits, the server address, and the model identifier. For local inference, also verify the configured context. Adding skills will not resolve an access error.

The agent seems to forget an instruction

Check that memory is enabled, that the information was saved, and that you are using the same profile. Then test a new session. If the information appeared only in a previously read document, it may never have been retained as a lasting fact.

A skill no longer produces the right result

Review the procedure and compare its prerequisites with the current environment. Identify the failing step before requesting a complete rewrite. Keep the error message and test the correction on your test data.

Spending increases without a usable result

Stop the task and examine repeated calls, tool errors, and searches that are too broad. Restate the objective with a specific deliverable and a stopping condition. Then run a limited trial instead of extending a loop that is already failing.

Checklist before regular use

An installation is ready when you can explain where it stores data, which service receives its requests, and what effects its tools can have.

• The installation came from the official domain, and any errors have been reviewed.
• The provider, model, and effective context capacity are identified.
• A simple task succeeds, and its session can be resumed.
• Persistence has been tested with harmless information.
• Selected skills have been reviewed and tested.
• Secrets are excluded from documents and shared material.
• File and network access match the task’s requirements.
• Actions with significant consequences require appropriate approval.
• Spending is tracked, and the limits are understood.
• A backup can restore the configuration and procedures.

After an update or a provider change, rerun your reference tasks before restoring automation. This checks how your installation behaves, beyond simply confirming that the program starts.

Related pages

• Hermes Agent
• OpenRouter
• Ollama

Frequently asked questions

Can I use Hermes Agent without a subscription?

You can use the software without subscribing to a commercial Hermes plan. You still need to provide access to a model through a separately billed API or an inference server you operate. External services used by its tools may also charge fees.

Does Hermes Agent learn by changing its model?

Memory and skill creation store information and procedures outside the model. On their own, they do not train its parameters. A saved procedure can therefore preserve a mistake, and you need a way to review and correct it.

Can I install Hermes Agent directly on Windows?

The official documentation consulted for this guide provides a native PowerShell installer and a desktop application. WSL2 offers another route using the Linux installation process. Check which features are available for your chosen installation method.

Does Hermes Agent keep its memory private?

Storing data on your computer does not guarantee that it stays there. When you use a remote API, memory included in requests is sent to the model provider. Assess storage, inference, and the services called by tools separately.

Is Hermes Agent suitable for unattended automation?

Start with a limited task whose output you can verify and whose effects you can reverse. Unattended operation also requires an access policy, spending limits, backups, and a process for handling failures. Actions with significant consequences still need an appropriate approval step.