Hermes Agent
Its memory and skill loop learns across sessions, running locally, remotely, or through messaging apps.
What it is
Hermes Agent is an open-source personal AI agent from Nous Research. You can hand it a task in the terminal; it calls the shell and external tools to carry it out and saves the task process as a session. When it runs into a reusable approach, it can also create a skill, and keep revising that skill when it is invoked later. The next time it handles a related task, it searches past sessions so that what was accumulated before takes part in the current run.
What actually sets it apart from an ordinary chat client is that it does not just return an answer: it links terminal operations, messaging entry points, long-term memory, and a skill system into a closed execution loop. The model source is not tied to the program either, and can be switched between Nous Portal, OpenRouter, OpenAI, or a self-hosted endpoint.
How to run it
Linux, macOS, WSL2, and Termux can run the install script. After installing, reload the shell and then start the interactive terminal:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.bashrc
hermesWindows has a native PowerShell installation path. On first use you can run hermes setup to configure the model and tools. To send tasks through Telegram, Discord, Slack, WhatsApp, or Signal, start hermes gateway.
Core mechanics
- Closed-loop learning: creates skills from the experience of carrying out complex tasks, and modifies the skills during later real invocations.
- Cross-session memory: retrieves past sessions with FTS5, then has the model summarize the relevant content, while also maintaining a user profile.
- Model decoupling: use
hermes modelto change the model service without modifying code, and you can also connect your own compatible endpoint. - Multi-backend execution: terminal tasks can run locally, in Docker, over SSH, or in Singularity, Modal, Daytona, or Vercel Sandbox; independent sub-agents can handle workflows in parallel.
Who it is for
It suits people who keep an agent deployed long-term on a personal server and want to submit tasks at any time through a messaging app, and also suits teams researching agent skill learning, trajectory generation, and tool-calling flows. The built-in timer can be used for unattended jobs such as daily reports, backups, and audits.
Boundaries
The project is MIT-licensed and can be self-deployed, but inference still usually requires a model service, a self-hosted inference endpoint, or the corresponding API key. Different execution backends and messaging platforms need to be configured separately. Termux uses a trimmed dependency set, because the full voice dependencies are not compatible with Android for now. Because the agent can execute shell commands and external tools, you should enable command approval, container isolation, and private-chat pairing for messaging when deploying.