hermes
Hermes is Nous Research's ambitious attempt at a personal agent that grows with you across every surface — CLI, Telegram, Discord, Slack, WhatsApp, email. Built-in cross-session memory, autonomous skill creation, scheduled tasks. MIT licensed. It's not a memory layer; it's an agent that has memory as a core feature.
- Technical users who want a fully-featured personal agent
- People who talk to assistants across multiple apps (not just Claude)
- Builders who want to see what a 'complete' agent stack looks like in OSS
What you'll do
Hermes is a full Python agent runtime. You install it, configure your LLM provider and surfaces, then run. Budget 30 minutes for the first setup.
Before you start
- Python 3.11+
- An LLM API key (OpenAI, Anthropic, or a self-hosted model)
- Optional: bot tokens for any messaging surface you want (Telegram, Discord, etc.)
Step-by-step install
- 011. Install
pip install hermes-agent
- 022. Initialize
Run the interactive init — it'll walk you through the first configuration.
hermes init
- 033. Configure your primary surface
Start with CLI. It's the simplest and lets you verify the agent works before wiring Telegram / Slack / etc.
- 044. Run the agent
hermes run
- 055. Layer on additional surfaces
Once CLI works, add one surface at a time. Each requires its own bot token/API setup. The hermes docs walk through each.
Your first 10 minutes
- 01Chat with it over CLI. Teach it who you are.
- 02Exit. Re-run. Confirm it remembers.
- 03Wire one additional surface (Telegram is easiest). Confirm memory persists across CLI and Telegram.
- 04Give it a scheduled task ('remind me about X at 9am tomorrow').
- 05Add Cognition CLO for your org's retention layer (hermes owns your personal memory; CLO owns team retention).
Troubleshooting
hermes init fails on Python version.
Hermes needs 3.11+. Use pyenv or conda to get a newer Python.
Telegram bot doesn't receive messages.
Confirm webhook/polling config in hermes config and that your bot is added to the chat. Telegram's BotFather docs are the source of truth.
hermes holds the knowledge. Cognition CLO models retention per employee per concept using a Weibull forgetting curve — so you see decay before it becomes a missed SOP or a failed audit.