The case for Hermes.

I've spent the last few weekends with Hermes Agent — Nous Research's open-source autonomous agent — and I keep going back to it after closing the laptop, which is the most useful signal I know for whether a tool is doing something genuinely new.

I came to Hermes after a stint with OpenClaw, which I also liked. On paper they look like cousins: open-source, self-hosted, bring-your-own-LLM, designed to live behind your messaging apps and act on your behalf. But they're actually betting on two different ideas about what a personal agent should be, and the difference shows up more the longer you use them.

The bet most agent harnesses are making

Most agent harnesses today — OpenClaw very much included — are designed to be excellent bridges. You ask the agent a thing, it figures out which tools to call, it does the thing, the conversation ends. The next time you come back, you're starting fresh. The agent has no real idea what your projects are, what your preferences are, what worked last time, or what you've asked it to do twelve times before.

That's fine for most use cases — actually great for them. Bridges are useful. But it caps what a personal agent can become. It stays a very capable calculator forever.

The bet Hermes is making

The Hermes README opens with a line that I think captures the whole project: "the agent that grows with you." That's the bet. What if the agent actually accumulated something? Not by retraining the underlying model — that's someone else's problem — but by building up its own structured memory of you, your work, and the skills it picks up along the way?

The pieces that make this concrete:

• Persistent memory with self-curation. The agent doesn't just log conversations — it nudges itself to retain things worth retaining, and it can search across past sessions with LLM-summarized recall. Your context isn't reset every time you say hello.
• Skills that auto-create from experience. When Hermes works through a complex task, it can save the procedure as a reusable skill. The next time you ask for something similar, it pulls from what it learned. The skills format is compatible with agentskills.io, an emerging open standard, so they're portable — you can share them, fork them, or pull in ones the community has built.
• A dialectic user model. Built on Honcho, a system designed specifically to build deepening profiles of users across sessions. Hermes's internal model of you isn't wiped on every conversation.
• Built-in scheduling. A cron-style scheduler lets you wire up unattended tasks — daily reports, weekly digests, nightly backups — and the agent runs them autonomously, then reports back via whichever messaging channel you've connected.

The detail that made me laugh

Buried in the Hermes docs is a command: hermes claw migrate.

It imports your settings, memories, skills, and API keys directly from an OpenClaw install. Nous is making it deliberately easy to defect from OpenClaw and bring everything with you. That's a confident move, and it tells you exactly how the team thinks about their position in the space. It's the kind of detail that says "we know what we're building, and we know why someone would switch."

What it actually feels like to use

A few sessions in, Hermes starts to feel less like a tool you summon and more like an environment you live in. The shift is subtle but real:

• Project context that sticks. Tell it once that you're building a personal site with Netlify Functions and a specific design system, and that context shapes every subsequent answer. You stop re-pasting your stack into every prompt.
• Workflow capture. Walk it through a multi-step task — pull data, filter it, summarize, post to Slack — and the next time you ask, it executes the captured skill instead of re-deriving the steps.
• Cross-session recall. You can reference work from a week ago without needing to re-summarize it. "That report we worked through on Tuesday" actually resolves to the right thing.
• One agent, every surface. The same Hermes instance answers from the CLI, Telegram, Slack, Discord, email — whatever you've wired up. Same memory, same skills, same model preferences, regardless of where the conversation happens.

Getting started

Hermes is genuinely easy to install. On macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.bashrc && hermes

That drops you into the configuration wizard. Connect a model provider (OpenAI, Anthropic, OpenRouter, Nous Portal, or one of 200+ options), decide which tools to enable, and you're in a working CLI session.

A few commands worth knowing right away:

• hermes — start the TUI
• hermes setup — full configuration wizard
• hermes gateway — start the messaging gateway (Telegram / Slack / Discord / etc.)
• hermes model — switch LLM provider on the fly, no code changes
• hermes tools — enable or disable specific tools
• hermes doctor — diagnose configuration issues

If you're already on OpenClaw: hermes claw migrate will pull everything across. Worth trying side by side for a week before deciding which one stays.

What to know going in

A few things I'd want someone to tell me before they sank a weekend into this:

• It's self-hosted, which means you're the ops team. "Runs anywhere" is genuine — a $5 VPS works fine — but you're responsible for keeping it running and updated. That's the trade for owning your data and paying nothing for the harness itself.
• Memory is a feature and a footgun. The fact that Hermes remembers everything is the whole point. It also means you should think about what you tell it the way you'd think about what you store in a password manager, not the way you'd think about a stateless chat. Back things up.
• You still pay your model provider. Hermes is free; the LLM behind it is not. Budget accordingly, especially with autonomous workflows that can chew through tokens unattended.
• It's moving fast. The repo is active and the feature set is shifting week to week. That's a feature if you like being early, a bug if you want stability. Pin a version and update deliberately.

Where this points

The frame I keep coming back to: most agent tools today are still selling you a smarter session. Hermes is selling you a smarter relationship. Whether that's worth the added complexity depends entirely on what you want a personal agent to be — but if the answer is "something that actually grows with me," it's the most coherent attempt at that I've seen.

Either way, the next year of agent tooling is going to be defined by this split: stateless harnesses optimizing for the perfect single turn vs. stateful agents optimizing for the long arc. Hermes is the clearest bet on the second answer, and it's worth a weekend just to feel the difference.