Hermes: The AI Agent That Keeps Getting Better at Its Job

You know that feeling when you explain something to a colleague once, and they just get it from then on? No repetition, no follow-up emails, no "per my previous message" energy?

That's Hermes. But it also runs on a $5 server and never asks for a raise.

What does Hermes even do?

Hermes is an AI agent built by Nous Research that comes with a built-in learning loop. The README calls it "self-improving," which sounds like marketing until you dig in and realize they mean it literally.

Here's what happens: you talk to Hermes, ask it to do things, and it watches. When it figures out a pattern—say, you always want meeting summaries in bullet points with action items highlighted—it creates a skill. A little piece of reusable capability that lives in its toolbox. Next time, it doesn't need the full explanation. It just does the thing.

The skills themselves improve during use. It searches its own conversation history to remember how you handled similar situations. It builds what they call "a deepening model of who you are across sessions."

Translated: it learns your preferences without you having to write them down.

Where It Lives

The install is one curl command. Works on Linux, macOS, and WSL2. Windows users need WSL2, which is mildly annoying but also the kind of friction that filters out people who probably shouldn't be running autonomous agents on their main machine anyway. Just switch to Linux its gotten way more user friendly in recent years, and with AI help you cant use “I don’t know shell” as an excuse.

Anyway, once it's installed, you've got options:

The CLI is where you'll probably start. Full terminal UI with multi-line editing, slash commands that autocomplete, conversation history you can search. You can interrupt it mid-task, redirect, watch tools run in real time. It feels like a proper development tool, not a chat bot bolted onto a terminal.

The gateway is where it gets interesting. Run hermes gateway and suddenly your agent is on Telegram, Discord, Slack, WhatsApp, Signal, and email. Voice memos work. Cross-platform conversation continuity works. Start something on your phone, finish it on your laptop, the agent remembers where you left off.

The infrastructure is deliberately flexible. Local machine, Docker container, SSH into a box somewhere, or serverless on Daytona or Modal. The serverless option means it hibernates when idle and wakes when you message it. Costs nearly nothing between sessions. You're not paying for a cloud VM to sit there waiting for you to remember it exists.

The Practical Stuff

Let's talk about what you'd actually use this for.

Scheduled Tasks

The cron scheduler is genuinely useful. You write in natural language: "send me a daily report every morning at 8am with yesterday's metrics" and it just... does that. Daily backups, weekly audits, monthly summaries—all running unattended, delivered to whatever platform you want.

If you've ever wrestled with cron syntax only to discover your backup script failed at 3am because of a typo, this feels like cheating. It is cheating (kinda). Enjoy it.

Parallel Work

Hermes can spawn sub-agents for parallel workstreams. You ask it to research something, draft a response, check your calendar, and summarize a document—it spins up separate agents for each, then collapses the results back into one conversation turn.

The technical term is "zero-context-cost turns" but the practical effect is you don't wait for one thing to finish before asking for the next. It's like having a team instead of an intern.

Skills and Memory

The skills system is where the self-improvement lives. Hermes creates skills automatically after complex tasks. You can browse them with /skills, use them directly with slash commands, or let the agent figure out when they're relevant.

The memory system uses something called FTS5 session search with LLM summarization. In practice, this means you can ask "what did we decide about the dashboard redesign last month?" and it actually knows. It's read its own diary.

The Migration Story

If you were using OpenClaw (the predecessor), Hermes imports everything. Memories, skills, API keys, personality files, the whole lot. The migration tool offers a dry run so you can preview what's coming across. It's thoughtful, careful, and exactly what you want when moving something that knows this much about you.

The Security Reality

Here's where the README gets honest in a way most projects don't.

Hermes can run commands on your system. That's the point. But you probably don't want it running rm -rf / because you asked it to "clean up some space."

The security documentation covers three main safeguards:

Command approval means it checks with you before doing anything destructive. You get to approve or reject before the command executes.

DM pairing verifies who's talking to it across different platforms. Someone can't just find your Telegram bot and start asking it to delete files.

Container isolation means when it does run things, it does so in a box that can't access the rest of your system. Docker, mostly. You can configure how strict this is.

The realistic threat isn't malicious behavior. It's slow drift. Today it asks permission. Over time, as it learns your patterns, it might start making assumptions about what you'd approve. The approval system is a safeguard against that drift, but only if you pay attention to what it's asking.

What It Doesn't Do

The README is refreshingly clear about limitations. Windows isn't supported natively. The CLI expects a Unix-like environment. The serverless options require accounts with Daytona or Modal. The gateway needs API keys for whatever platforms you're connecting.

It's not a magic box that solves everything. It's a tool that does specific things well, and the documentation tells you exactly what those things are.

The Installation Walkthrough

If you're going to try it, here's the flow:

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

The setup wizard walks you through choosing a model provider (OpenRouter, OpenAI, local endpoints, etc.), enabling tools, and configuring memory. It takes maybe five minutes.

Then hermes starts the CLI. /model switches providers if you want to experiment. /personality sets tone and style. /skills shows you what it's learned.

For the gateway: hermes gateway setup then hermes gateway start. Follow the platform-specific instructions for Telegram, Discord, etc. Each one needs a bot token and some configuration, but the wizard handles most of it.

The Thing Worth Watching

Hermes is described as "research-ready" with "batch trajectory generation" and "RL environments" for training future models. This means every conversation you have, every task you delegate, is potentially training data for the next version.

Not for some corporate AI somewhere. For your agent. The one that lives in your infrastructure and remembers how you like things.

That's the loop worth paying attention to. It's learning you. And the more it learns, the more useful it becomes. Also the more it starts anticipating instead of just responding.

Should You Run It

If you spend any time in the terminal, if you have recurring tasks you'd rather not think about, if you've ever wished your tools would just remember how you work—yes, it's worth spinning up.

The $5 VPS option is genuinely cheap. The serverless option is cheaper. The local install is free. The time it saves on scheduling alone pays for itself inside a week.

Just read the security docs. Set the approval thresholds where you're comfortable. And maybe check in on what skills it's creating every once in a while. The self-improvement loop is the feature. It's also worth keeping an eye on.