3 months with Hermes Agent: what i wish i knew earlier

i just crossed three months using Hermes Agent as my daily driver.

three months with the software, and three months inside the community around it. i started on an old phone, mostly out of curiosity. at the time, almost nobody seemed ready to trust this “new tool” seriously. after a few days, though, i realized Hermes was becoming more than something i wanted to test.

so i bought more appropriate hardware just for it.

about four weeks in, around the time i wrote my “one month with Hermes Agent” post, i stopped thinking of Hermes as just a tool. because it became magical? the opposite, actually. i started understanding it better. i began seeing where its answers came from, what context it was using, what an healty config should looks like, why some outputs worked, and why others failed.

my biggest lesson from three months with Hermes: the more you understand it, the less magical it feels. then, weirdly, that is where the actual magic begins.

The elephant in the room

memory.

a lot of people come to Hermes because they hear about “the agent that remembers“. sounds great, and in some ways it is. but after three months, i think the expectation is usually wrong.

i tested every memory provider connected to Hermes. i also tried a few open-source ones trending here and on GitHub. eventually, i turned them all off. for me, memory introduces trade-offs very quickly. bad memory is worse than no memory. unclear memory creates weird outputs. too much memory makes the agent feel like it knows things, when in reality it is just pulling from messy context and messing with your work. am i using Hermes the same way i was 3 months ago? hell nah. and i don’t want a layer to hopefully pick that change and inject it at the right time, without me really knowing.

agents do not remember. agents read.

Hermes already comes with a md USER / MEMORY / SOUL system. after trying more complex setups, i honestly think that default structure is probably the best starting point. guess what? whoever built this agent probably already knew this. hence the choice.

USER and MEMORY are short and editable by the agent (just like skills are). SOUL stays immutable. given that, my current advice is simple: leave that system in place, inspect it regularly, clean it when something wrong gets in, and tweak it on purpose when you want to see how the agent changes (it’s fun, it’s like manipulating someone’s brain, and it teaches you a lot). but do not assume more memory automatically means a better agent.

for me, it usually meant more noise. and more frustration when pollution entered my workflow. you can feel that.

Six profiles became three

another thing i changed: profiles.

i used to argue that profiles are one of the foundations of Hermes, and i still believe that. choosing which tools, skills, and context each agent has is basically half the job.

but i also overdid it.

at one point, i had six profiles: one default, one for code, one researcher, one for my girlfriend, one archiver, and one for experiments.

three months in, i use three profiles. they are scoped hard, i’m leaning towards more procedural skills, and agents with the right set of tools and instructions to deal with all of them. check my link-curator profile for reference.

another thing i started doing a lot now, is take my time to create and seed good docs that automatically turns into the right context when my agent interact with a project. private repos have been great for that. also for letting agents share docs they wrote, and so their context.

if you have more than four profiles, good chance you probably have more than you need.

My terminal got less scary

this is one of the biggest practical changes in my workflow.

before Hermes, a lot of tasks had to go through raw commands. now, the first question is usually: can Hermes do this?

sometimes the answer is no, and i do it myself (or i just complete the step it can’t help with). but most of the time, the answer is: describe the task well and let it surprise you.

GIFI’ve often found myself smiling at my phone these past few weeks, with my girlfriend asking me why…

I’ve often found myself smiling at my phone these past few weeks, with my girlfriend asking me why…

starting a task is almost free now (time speaking). i am thinking about the task itself, it shifted to describing. i now spend more energy defining what i want and less energy fighting the interface between me and my machine. exactly what i wanted when i started with Hermes, just in a totally different way than i thought it would.

The hardware part still feels ridiculous

my Hermes swarm runs on a $50 Raspberry Pi 4 with 8GB of RAM.

i got the agent for free, and it is running on hardware that costs less than dinner. you pay for what you use, and you decide where your money goes (!). or, if you are skilled and equipped enough, you run everything locally and let the world burn around you without caring much.

“that update broke my setup”, never happened here. no mysterious regression that forced me to rebuild everything. just running, the same agent becoming the primary interface to my own machine. and i repeat: i got it for free. you find rough edges? well, ask for a refund then.

It has not blinked.

So, when do you “master” Hermes?

I don’t know.

after three months, i am not sure “mastering” it is even the right framing. Hermes changed the speed at which i learn other things. and forced me to learn a lot, but i just scraped the surface.

i started by treating it like a new toy. i gave it a room. then i tried to make it smarter, louder, more complex. more memory. more profiles. more experiments. at some point, the opposite became true: the cleaner the setup, the better Hermes worked.

for me, Hermes works when you understand what you are giving it. context, tools, memory, profiles, configs, all of it. the deeper you go, the more it becomes yours. even tiny things, like discovering there is a built-in notification setting for the TUI

sorry, no “how i made 10k a month using Hermes Agent” story here. no magic trick, i’m not a magician probably. just a small open-source agent running on cheap hardware, slowly becoming the main way i interact with my own machine.

that is where Hermes is for me right now.

at least from where i’m standing, is redefining how humans interact with technology. and i think that is actually magical.

thanks to @NousResearch, @Teknium, the people running the community, and everyone shipping this thing daily.

hoping for an even brighter future.

🦤