A self-hosted agent from Nous Research that remembers, schedules work, and writes its own skills.
Aki Wijesundara, PhD
Instructor
This session is about what Hermes is and how it learns. Setup stays at the end as a short starting path.
An autonomous agent that compounds over time
Memory, skills, recall, and user modelling
CLI, messaging, cron, and voice from one agent
SKILL.md playbooks that write and improve themselves
Tools, MCP, delegation, and scheduled work
How the two projects make different bets
Recurring work that deserves muscle memory
One install path and one first skill
Hermes Agent is an autonomous agent that gets more capable the longer it runs. It is built by Nous Research and works with any OpenAI-compatible endpoint.
Persistent memory carries preferences, environment facts, and lessons across sessions.
Successful workflows become SKILL.md playbooks that Hermes can reuse later.
Talk to it from the CLI or messaging apps while it runs on a VPS, laptop, or serverless host.
Hermes creates skills from experience, improves them during use, persists knowledge, and builds a deepening model of who you are across sessions.
Hermes is organised around a learning loop, not around a single chat window. Read the diagram clockwise from 01.
Hermes solves a task with tools, the same way any agent would in a single session.
Useful facts land in MEMORY.md and USER.md, ready for the next session start.
The next similar task loads skills and memory instead of rediscovering the approach from scratch.
A background review can write or patch a SKILL.md so the workflow becomes reusable.
The same Hermes runtime powers the CLI, messaging platforms, scheduled jobs, and voice. Front ends change. The agent and its memory stay.
Core conversation loop, tools, and prompt assembly
Telegram, Slack, Discord, WhatsApp, and more
Scheduled agent jobs delivered to any platform
Real-time voice in CLI and supported chat apps
Hermes splits knowledge by cost. Small facts stay in every prompt. Longer procedures load only when needed. Full history is searchable on demand.
Small durable facts stay in every session prompt. They live in ~/.hermes/memories/ and are injected at session start.
Longer playbooks live under ~/.hermes/skills/. Hermes loads a skill only when the task matches its description.
Past CLI and messaging sessions stay in SQLite. Hermes can retrieve real messages without putting the full history into every prompt.
Honcho, Mem0, and similar plugins can deepen user modelling beside this stack. They do not replace the built-in files.
A skill is an on-demand instruction document. Hermes reads short descriptions for free, then loads the full SKILL.md body only when the task needs it.
Each skill has a name, a short description, a step-by-step procedure, and optional scripts or reference files.
Adding skills does not bloat every request, because the full playbook is loaded only when relevant.
Skills follow agentskills.io, so they are portable across tools and community hubs.
If you already write SKILL.md files for Claude Code, you already know the muscle Hermes uses.
Hermes can solve a problem once, save the working approach as a skill, and reuse that playbook the next time a similar task appears.
Skills become slash commands. You can also author with /learn. Setting skills.write_approval to true stages writes until you approve them.
A background maintenance pass archives unused agent-created skills so the library does not fill with near-duplicates. Hub skills stay untouched.
Beyond memory and skills, Hermes ships as a full agent runtime with tools, scheduling, and parallel work.
Built-in terminal, browser, and file tools. MCP servers extend the set safely.
Cron runs agent jobs on a schedule and can deliver results to any connected platform.
delegate_task spawns isolated children, up to three in parallel by default, then returns summaries.
Local machine, Docker, SSH, or serverless hosts such as Daytona and Modal when idle cost matters.
Nous Portal, OpenRouter, OpenAI, Ollama, or any OpenAI-compatible endpoint.
Batch runs and trajectory export support training workflows from the same agent runtime.
Both are self-hosted agent systems. They make different bets about what should improve over time.
| Dimension | Hermes | OpenClaw |
|---|---|---|
| Core bet | One agent that improves through memory and skills | An organization of agents under a control plane |
| Shape | Python CLI first, with an optional messaging gateway | A central gateway daemon with hub-and-spoke routing |
| Multi-agent | Temporary subagents through delegate_task | Persistent named agents bound to channels |
| Skills | Agent-written and hub skills via agentskills.io | Large built-in set plus marketplace plugins |
Choose Hermes when automation should get better over time. Choose OpenClaw when the main problem is orchestrating many agents across channels.
Hermes fits recurring workflows that happen often enough to become skills and improve with use.
Install is intentionally short. The compounding value starts after Hermes has a model and one real recurring task.
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashRun hermes model, or hermes setup --portal for Nous Portal OAuth and tool gateway access.
Start in the CLI, then optionally connect Telegram or another platform with hermes gateway setup.
Write or let Hermes write one SKILL.md for a workflow that already repeats every week.
Secrets belong in ~/.hermes/.env. Behaviour settings belong in config.yaml. Docs: hermes-agent.nousresearch.com/docs
Hermes is valuable because the loop keeps running after the first chat.
Memory, skills, recall, and optional user modelling feed each other.
CLI, messaging, cron, and voice share the same memory and skill library.
Install, connect a model, then capture one recurring workflow as a skill.