Hong Cheung 
How I Set Up Hermes Agent
What I did was having it running on a VPS server that I self-hosted via Hetzner1. Using Telegram as the gateway to talk to my GPT on my phone. And I was using OpenRouter Owl Alpha model, which was free at the time and it has 1M token limit. It is quite enough for anything.
Long-Term Memory and References
I have also setup MCP-Memory for saving memories locally on the VPS, so I don’t need to repeat myself every time I start a new session. And MCP-Local-Rag, for local books collection I had to give reference to my questions. It covers topics like business, health, science, etc.
Best Use Case of Hermes Agent (for Me)
Learning Languages
I found the best use case for me was learning languages with it. The way it works was I created a language learning agent named Fiona. And for every session, when I start with “Hey Fiona” it will local all the language learning skills and context into the session, so we know what we were about to do.
And for the language parts, I’ve been learning Spanish and Japanese. The language learning in this workflow are mostly just vocabs to meanings kind of quizzes I’ve set up with my GPT.
Finding The Vocabs for My Targeted Language
I asked it to go to github to find and download csv/md files for both languages, and categorise them inside the local folder /vocabs/Spanish, etc. I would name those files N1-N5 and A1-C2 depending on their levels. So I would have N1.csv, A1.md - it was a mix of filetypes.
Quizes
I setup a rule with Fiona to ask me the meanings of my targeted language 5 words at a time, and for every failed attempt, it will got written down how many times I got this word wrong into another wrong-N1.csv file. For the right one, I modify the original file and add 1 score to that row. The higher the score, the less likely it appears in the quizzes. So, the quiz words are random yet optimised for me.
Reminded Vocabs Challenge
This is something I found useful. I’ve setup some cronjobs to have it automatically send me quizzes throughout the day everyday. Normally I would get two quizzes send to me at around 8 a.m. where I have my break from work. Another two at 11 a.m. right before the lunch break. And another two before bed, which was ideally 8 p.m.
Overall Experience on Using It for Language Learning
I’d say it’s a mix. It could be that I didn’t set it up precisely enough, or that relying on LLM to do things on it’s own repeatedly is just not stable enough. Apart from sometimes decent “work as I told” experience with it. It had ran into hallucination and off tracked to run all sorts of weird bash every now and then. Which makes the whole experience not as pleasant.
Other Use Cases
I’ve tried to use it just for chatting. You know, having a personal assistant that remembers my everything and tailor the answer just for me, with all my ambitions, habits, and situation considered. But in the end, it can never replace a human for a valuable chat. You don’t know what you don’t know. I can set it up the way I wanted, and I believe that’s the downside.
It’s like people trying to use GPT as AI partner, they can set it up the way they wanted. Which is too ideal and it’s missing the core human elements such as mood, their own vision, their feeling about you, connections between people, etc. So, talking to a GPT like a friend, really won’t work nicely. I don’t need an echo chamber or a forever comforter. I want to hear experiences in real life.
That being said, the only way I found GPT or AI that’s kind of useful, is to just using it as search engine. That’s pretty much it. No more ads on webpages, clicking through hoops. And I prefer text to start with.
Alternative for Hermes Agent (Not OpenClaw)
It is interesting that I found using OpenCode, or any programming IDE - the tools that are made for AI programming are more useful than Hermes Agent in general.
The main difference between them and Hermes Agent is that, most of them are project-scoped. Meaning they can only work inside that folder, for each project you’ve declared. But on the other hand, by default, to leverage the full power of Hermes Agent / OpenClaw, you would usually give full permission to them, which means they can, and will, mess up your other files and settings in your machine.
Programming IDE, a new folder per agent is how I work now. I now have language learning agent just work inside that folder, and investing researcher and advisor just inside another folder. Each of them have different set of rules and guardrails. I can easily switch between them to work on different subjects and make adjustment to the files itself as oppose to using Hermes Agent.
Conclusion
Hermes Agent sounds innovative, and it might be. But I kind of feel like it is over engineered to serve a not quite necessary itch. Because at one point when I was using it for language learning, I was thinking, heck, at this point I might just build my own flashcard app and it would have been more resource light, and most importantly very stable.
But who knows, maybe one day Hermes Agent will be running in a robot, and Hermes Agent become a robot OS. Considering how big of a push it gets from Nvidia for releasing their new AI computers with Hermes Agent built-in.
But for now, I’ve already uninstalled it and kept using OpenCode and KiloCode for my daily usage, and it has been good.
Oh, I also started using paid models because of privacy concern as well.
