How to use Hermes Agent: train an AI that never forgets

Using Hermes Agent well is a different sport from using a chatbot well. A chatbot session is a transaction; a Hermes deployment is a relationship. The agent writes itself skills after hard tasks, remembers everything worth remembering, and builds a model of who you are, which means how you use it in the first weeks decides how good it eventually gets. We run Hermes agents around the clock at PlusAgents, and this is the field guide: how to start, what to delegate, and how to make the learning loop compound. New to the project? Our What is Hermes Agent explainer covers the background.

Step 0: get it running (and keep it running)

One thing to internalize before anything else: Hermes is designed to run 24/7 on a server. Its memory, skills and scheduled jobs only compound if the agent is actually alive between your conversations. A laptop that sleeps every night is the one deployment Hermes does not love. Your options:

• One click on PlusAgents. We deploy a private, always-on Hermes instance with LLM credits included. Free to start, live in about a minute.
• Self-host. One curl command on any Linux box, Mac or WSL2, famously happy on a $5 VPS. Full walkthrough in our Hermes deployment guide.

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes setup   # connect a model provider
hermes         # start chatting

The first conversation: introduce yourself properly

Your first session sets the foundation for the user model, so do not waste it on "tell me a joke". Spend ten minutes actually briefing your new hire:

• Who you are, what you do, and the projects you are juggling right now.
• How you like output: brief or thorough, bullets or prose, direct or diplomatic.
• Your stack and your world: the tools you use, the people you mention often, the timezone you live in.
• What you plan to delegate, so it knows what to pay attention to.

This feels odd with software. Do it anyway. Hermes persists what it learns, nudges itself to save important context, and searches its own history later. Every minute of briefing now saves ten minutes of re-explaining forever. Then give it a first real task, ideally from the terminal so you can watch the streaming tool output: research something messy, summarize a repo, draft a document.

Week one: grow the skill library on purpose

Here is the mental shift that separates power users from tourists: with Hermes, hard tasks are investments. When the agent grinds through something complex (a gnarly data cleanup, a multi-step deployment, a research workflow with your specific sources), it writes itself a skill document capturing the how. Next time, it does not rediscover the process; it executes it.

So in week one, deliberately feed it the recurring shapes of your work:

1. Pick three tasks you do every week and hand each one over once, with patient corrections.
2. When it gets one right, say so explicitly. Outcome feedback sharpens the skill it writes.
3. Ask it to show you its skills now and then. You can read, prune and improve the library, and it is compatible with the agentskills.io standard, so community skills can be imported too.

By week three, watching your agent handle a formerly painful task in one clean pass, using a skill it wrote itself, is the moment most people become believers.

Memory: trust it, search it, prune it

Hermes remembers across sessions by design: agent-curated memory plus full-text search over every past conversation. In practice that means you should talk to it like a colleague with perfect recall. "What did we decide about the pricing page in June?" and "use the same format as last month's report" are normal requests. If it ever saves a wrong impression (it happens), just tell it; memory is editable, and correcting early keeps the user model clean.

Schedule it: the compounding really shows here

The built-in cron scheduler is where an always-on agent earns its keep. Ask in plain language:

• "Every weekday at 7am, send me a briefing on Telegram: calendar, inbox highlights, and anything new on the topics I follow."
• "Every Friday at 4pm, draft my weekly status update from what we worked on this week."
• "Every night at 2am, back up the project folder and tell me only if something fails."

Since the Pantheon release (v0.21.0), scheduled jobs have memory and continuity between runs, so your Friday report writer actually gets better at writing Friday reports. Delivery goes to any connected platform.

Delegate bigger jobs: subagents

For work too big for one conversation, Hermes spawns isolated subagents, each with its own conversation and terminal, running in parallel. "Research these five competitors, one workstream each, and merge the findings into a comparison" is a single request. Power users go further and have Hermes write Python scripts that call its own tools over RPC, turning a ten-step pipeline into one repeatable command. You do not need this in week one; it is nice to know the ceiling is high.

Live where you live: the gateway

Connect Telegram, Discord, Slack, WhatsApp or Signal (one gateway process serves all of them) and your agent follows you across platforms mid-conversation: start a thread on Telegram at lunch, continue in the terminal at your desk. It transcribes voice memos, so you can brief it hands-free. On PlusAgents deployments the gateway is already running; connecting a chat app is a guided flow in your dashboard.

A realistic four-week arc

• Week 1: brief it, connect one chat app, delegate three recurring tasks with corrections.
• Week 2: schedule the morning briefing and one weekly report. Review the skill library once.
• Week 3: hand over a genuinely hard multi-step job. Watch it write the skill.
• Week 4: notice you have stopped explaining context. That is the loop closing.

If Hermes sounds like more commitment than you want, the more instantly-gratifying OpenClaw might be your speed, and our honest comparison will help you decide. But if the idea of an assistant that is measurably better in month two than in week one appeals to you, there is no substitute.