Applied Hermes

Hermes Agent by Nous Research · Independent guide

Hermes Agent learns
how you work, then
keeps getting better.

Nous Research’s open-source agent remembers across sessions, turns the workflows you repeat into skills it can run again, and sharpens them as you go.

Independent guide. Updated for the current Hermes Agent release.

Official, from Nous ResearchProduct siteGitHubDocs
~/.hermesWhat survives the session
memories/MEMORY.mdFacts about your setup and projects, loaded fresh at the start of every session.cap 2,200
memories/USER.mdHow you like work done. The agent writes it; you can edit or delete it anytime.cap 1,375
skills/…/SKILL.mdA procedure it wrote down once and can load again on a matching task.reusable
state.dbEvery session you have had, searchable down to the message.FTS5 search

Local and inspectable. Memory and skills are stored as files you control; past sessions live in the local state.db database and remain searchable.

No Hermes account to create. You bring the model provider.

MIT licensedLaptop or always-on serverMultiple model providers

What are you here to do?

Hermes DesktopTerminalTelegram · Slack · email
~/.hermesmemory, skills, every past session

An open-source agent that keeps what it learns.

Hermes runs on your own machine, uses the model you choose, and reaches your files, terminal and browser. The part that makes it different is what happens after the task: it writes down what it learned and can run it again.

Built by Nous Research, MIT licensed

Desktop app, terminal, or chat platforms like Slack and Telegram

Memory, skills and past sessions live in ~/.hermes

Read the plain explanation

The learning loop

Session one is the worst Hermes will ever be.

Most agents meet you again from scratch every time. Hermes ends a session by writing down what it learned, how you like the work done, and any workflow worth running again, then starts the next one from there. That is stored context and reusable procedures carrying forward, not the underlying model retraining itself.

Illustration — how memory and skills accumulate across five sessions

You ask

Pull Thursday's revenue out of Postgres and write me a summary.

A cold start. Everything gets explained — connection, schema, what a useful summary looks like.

Carried in (0)

Nothing yet. First session, so everything has to be explained.

Written this session (3)

MEMORY.md · warehouse schemaMEMORY.md · summarise, don't dump tablesUSER.md · numbers first, prose after

In practice

What people actually point it at.

None of this needs a new interface. It is the same agent whether you reach it from the desktop app, a terminal or a chat thread.

See what it can do
01

Work that runs long

One thread that moves between the terminal, your files, a browser and web search without you re-briefing it at every handover.

terminal · files · browser · web search

02

Research that stays open

Sources, findings and the questions you have not answered yet survive the session. Weeks later it can search the actual messages, not a summary of them.

state.db · FTS5 search · MEMORY.md

03

The Thursday jobs

Recurring reports, checks and briefings described in plain language, run on a schedule and delivered where you already read things.

cron · gateway · Slack, Telegram, email

04

The thing you do every month

Once a workflow works, Hermes writes it down as a skill and runs it again. Point it at your own docs and it writes the skill from those.

skills/ · /learn · agentskills.io

On your machine

Your laptop is enough to find out.

One command installs it, and you approve commands as it runs them. Your config, memory, skills and session history sit in ~/.hermes where you can read them, and hermes uninstall removes the agent while leaving that data in place — add --full when you want the data gone too.

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

CLI on macOS, Linux, WSL2 and native Windows

Desktop app or terminal, same agent underneath

You pay for model usage, not for Hermes

The catch: it works when your machine is awake and you are there. Nothing runs while the lid is shut.

Always-on hosting

Scheduled work needs a machine that stays up.

The messaging gateway is the part that has to keep running. It holds your chat platforms open, answers from wherever you are, and keeps the cron scheduler firing whether or not you are logged in. On a server, that is a service. On a laptop, it is a coin flip.

$ hermes gateway install
$ sudo loginctl enable-linger $USER

systemd on Linux, launchd on macOS, restarts on boot

Reach the same memory and skills from your phone

Container backends when you want the blast radius small

A server you own is a server you secure. Non-root user, explicit allowlist, keys in ~/.hermes/.env at 0600.

First run

Four moves between nothing and an agent that knows your setup.

Most of the setup time goes into the first two steps. The fourth is optional, but it is the fastest way to see what Hermes kept from a session and what it did not.

01

Install it

One shell command on macOS, Linux or WSL2, or the PowerShell installer on native Windows. The desktop app has its own installer if you would rather not use a terminal.

curl -fsSL …/install.sh | bash
02

Point it at a model

Nous Portal, OpenRouter, OpenAI or any endpoint you already pay for. Switch later without losing anything.

hermes model
03

Give it real work

Not a demo prompt. Something you were going to do today anyway, so you can judge the result honestly.

hermes
04

See what it kept

The journey view lists the memory notes and skills written during your sessions, so you can check them and delete anything wrong.

/journey
Set up Hermes Agent

Step by step, with the parts that usually go wrong called out.

Hermes gets better by working with you. Give it something real to start with.

Hermes Agent is built by Nous Research. Applied Hermes is an independent guide to evaluating, installing and running it.