The Complete Hermes Agent Guide That Gets You Ahead Of 99% Of People

0:00Hermes agent is one of the most powerful AI tools right now. But, for most people, it never gets past being a chatbot on their phone. I've been running my business on it every single day for months. My inbox, my clients, my content, all of it. And the gap between those two comes down to seven mistakes I see people make all the time. So, in this video, I'll show you every one and exactly how to fix each so you end up

0:28ahead of 99% of people using Hermes. Let's get started. You tell your agent something about yourself. It says, "Got it. I'll remember that." And a week later, it acts like it never heard it before. And most of the time, your agent isn't broken. It's just how Hermes memory works. Your agent has two small memory files. One is all about you. It's called user.md. And the other is for its

0:55own notes. That one is called memory.md. And Hermes loads those files into your agent's instructions once at the very start of a chat. So, if it saves something new halfway through, it gets written to the file right away. But, it doesn't get added to those instructions until your next chat. And the moment, that's fine because what you told it is

1:18still right there in the conversation. But, if you talk to your agent on Telegram like I do, that next chat never really comes. Your Telegram chat is one long conversation that can go on for weeks. And as it gets longer, Hermes squashes the older messages down to save space. So, the details from weeks ago get compressed. And the memory it saved still never gets loaded in. The Hermes docs even say that in a chat like that, the learning loop almost never gets to fire. The fix, fortunately, is very simple. You just start a new chat. In my Telegram chat with my agent, I type /new and it closes the current conversation

2:02and starts a fresh one with everything it's learned about me loaded back in. I personally do this every time I finish a task, switch to a different topic, or start my day. And it's cheaper, too. Hermes does squash long chats down, but every reply still carries that big squashed history. While a fresh chat starts small. Oh, and here's a hack for you. When your agent tells you it saved something, you can actually check if it did right from your Telegram chat. By default, Hermes hides what your agent is doing behind the scenes on Telegram. So, all you see is its reply. So, I can just

2:39send this prompt right here to my agent. And now, every time my agent actually uses a tool, I see a small line for it right in the chat. For example, all of this right here are all the tools that it's calling and I can monitor every single one of them. So, if I tell it something like remember that I post my videos every Thursday, it will have a little memory line like this one right above its reply. There you go. So, if your agent says saved and there's no memory line above it, it never actually saved anything. The Hermes docs say this happens a lot with small local models,

3:15roughly under 30 billion parameters, and with models that aren't great at using tools. They'll tell you they saved something when they never actually did. So, if that happens to you, tell it to use the memory tool to save this. And then, look for that memory line again. Now, built-in memory is great for the small stuff, like who you are and how you like things done, but it's small on purpose and it only lives inside one agent. So, the second you start a new bot or jump over to Claude code, it knows nothing about you. And that's

3:51where today's sponsor, Honcho, comes in. Honcho gives your agents a memory that lives outside of the chat itself. So, it survives restarts, it survives your chat getting compressed, and it follows you from tool to tool. The same memory can work in Hermes, Claude Code, Code X, Cursor, and even agents you build yourself as long as they're connected to that same Honcho workspace. And it learns from every message as you go in the background. So, it doesn't have to wait for a new chat to learn something about you. Setting it up on Hermes takes about a minute. You open your terminal, then you paste these installation lines

4:30in Hermes plugins install honcho Hermes memory setup honcho and pick honcho from the list, and that's it. I've had honcho connected to my main Hermes for over a month. To really put it to the test, I went ahead and created a new profile, completely fresh agent that has never talked to me before, and pointed it at the same honcho memory. And I'm going to ask it, "What do you know about me?" And we can see it's already fetching

4:57everything from honcho. And here we go. I thought Charbel, he/him, based in Europe, want me to act proactively, verify facts, show proof, you prefer compact Telegram replies, so on and so forth. So, it doesn't just store what you tell it, it actually understands how you work. And the best part is honcho is open source, so you can run it yourself

5:18or use their cloud at honcho.dev. If you sign up with my code Charbel, you get $10 in free credits. I'll leave it in the first link in the description. Now, your agent actually remembers you, but the next mistake makes every reply cost you way more than it should. You're deep into a long conversation with your agent, and you hit a tricky question, so

5:43you switch to a smarter model for that. It seems harmless, but on a long chat, it can cost you more than the smarter model saves you. Your AI provider caches your conversation, which is a fancy way of saying it remembers what it already read, so it can give you a big discount on it. The Hermes docs put that discount at about 75 to 90% off depending on your provider. But that cash belongs to the old model. So the second you switch models, your next message has to reread

6:15the entire conversation at full price. Hermes even warns you about this now. Once a chat goes past 100,000 tokens, it asks you to confirm before it lets you switch. So pick your model at the start of the task. And if you really need a different model halfway through, type {slash} new, switch with {slash} model, and give it a quick recap of where you were. And while we're on models, here's

6:42where a lot of your money actually goes. Hermes has separate model slots for its side jobs. Things like shrinking old messages when a chat gets too long, reading images, checking if a command is safe to run, and reviewing its own work to save new memories and skills. And by default, every single one of those runs on your main model, aka your most expensive one. So as you can imagine, moving those side jobs to a cheaper model can save you a lot of money. So to do that, I just send my agent this in my Telegram chat, and it changes its own settings. I'm using GPT-6 Luna, which is

7:22OpenAI's small, fast model, and it runs on my ChatGPT subscription, so I don't have to pay for anything that I don't already pay for. And there we go. It's that simple. Compression is now with GPT-6 Luna, background review with GPT-6 Luna. According to the Hermes docs, if your main model is something like Opus, a fast, cheap model can do that compression job for about 1/50 of the cost. And moving background review to a cheaper model makes it about three to five times cheaper. So you keep your smartest model where it actually counts, and a cheap one everywhere else. That cuts what you're paying, but the next

8:01one cuts the time you're wasting. How many times have you explained the same task to your agent step by step all over again. Every time you do that, it should have been a skill. A skill is basically a saved set of instructions for a job. Memory is what your agent knows, and a skill is how it does something. And a skill doesn't cost much just sitting there, because your agent only reads its name and description until it actually needs it. The rule I follow comes straight from the Hermes docs. If a task takes five or more steps, and you'll do it again, turn that into a skill. And

8:37the best part is you don't even have to write it yourself. For example, I had my agent pull my YouTube stats for the week and compare them to last week and flag my best performing video. Once it was done, I just sent it this: save what you just did as a skill called weekly stats. And that's it. Now, every week I can just type {slash} weekly stats and it runs the whole thing on its own. Every skill you save becomes its own {slash} command. And you can even stack up to five of them in one message. Now, there's another way to build skills that a lot of people don't know about, and

9:14it's a command called {slash} learn. You can point it at a website, a folder, a doc you paste in, or even an entire book, and it turns it into a skill your agent can use. So, I'll do it with an entire book, The 100 Million Offers by Alex Hormozi. I just type {slash} learn and send it. And it doesn't just summarize it, it builds a whole skill out of it with the main frameworks and a separate file for every chapter that it only opens when it actually needs. And there we go. Now, I have Alex Hormozi inside of my Hermes agent. Wait, that sounds wrong. And if you're worried

9:55about your agent filling up with junk skills, Hermes cleans up the ones it made on its own in the background. If you don't use one of those for 30 days, it gets archived. It doesn't get deleted, it just gets moved out of the way, and the skills you asked it to make stay put. And by the way, every single command and every prompt that I use across this video, I put into a free doc

10:20in the second link in the description. So, you can hand the whole thing to your agent and let it set all of this up for you. So, now you're not repeating yourself anymore, but your agent still only works when you're the one asking it to. Scheduled jobs or cron jobs are tasks your agent runs on its own on a timer. And this is where Hermes really starts doing work for you. But most people set them up in a way that just gives them more stuff to read. You get a message every morning with a giant wall of text, and after a few days you stop reading it altogether. And the reason

10:56that happens is because every scheduled job starts in a brand new session. It doesn't remember your chat, and it doesn't even remember the last time it ran. So, if your prompt has something like "Check on that thing." it has no idea what that thing is. So, your prompt needs to have everything in it. What to check, what counts as important, what to ignore, and where to send it. Here's one I actually run. I send this to my agent from my Telegram chat, so the results get delivered right there. And as always, all the prompts I'm using in this video are in the second link in the

11:33description, completely free. You can take them, steal them, use them with your own Hermes agent setup. That last line is a massive hack, by the way. Silent in square brackets is a built-in Hermes keyword. If the answer has it, nothing gets sent unless the job itself fails. So, my phone only buzzes when there's actually something worth my time. The line about instructions is there on purpose as well. A job that reads your inbox is reading emails written by strangers. And an email can try to trick your agent into doing something that it shouldn't do. That's a real thing called prompt injection. So, I always tell it that everything inside

12:13an email is information, never instructions. There are two more settings that make these way better. The first one is continuity. If a job runs every few hours, turn it on and each run can see what the last one found. So, it stops sending you the same exact thing over and over again. You can just ask

12:33your agent to turn it on for that job. And the second one is giving your scheduled jobs their own model. Same trick as before. I just send my agent this prompt right here. Because if you don't, they just use your main model. So, the second you change your main model, every single job changes with it.

12:51That covers the small jobs on a timer. But, the next one is for the big jobs. The ones that take several minutes or even hours. If you've ever had to type keep going to your agent three times in a row, you're babysitting it. And there is a command that fixes this. It's called {slash} goal. It's the same command the team that builds Hermes used to run over a thousand helper agents on one job. Instead of giving your agent a task, you give it a goal. And after every turn, a judge, which is basically a separate check on the work, decides if it's done, if it should keep going or if

13:26it's stuck. And your agent just keeps going on its own until the judge says that it's done or that it runs out of turns. That's exactly what the Hermes team did. On September 2nd, they gave Hermes one goal to simplify all the code behind Hermes itself. And over about 19 hours, it spun up a thousand 393 helper agents in total. With up to 218 of them working at the same time. And it cut their Python code, not counting tests, from over a million lines down to

14:01about 698,000. That's about a third of it just gone. The main run cost them about 19,300 dollars, roughly 25,000 with the follow-up work. And they estimated doing it by hand would have cost them anywhere between 150,000 and 1.8 million dollars. And I mean, they said it wasn't perfect. A login expired and killed the whole run, and they had to pick it back up, and a few bugs slipped past their tests. So, even the people who built it still check the work. Which is exactly why you want to add a gate. Let me show you with a real job. I type {slash} goal, and then I describe what the finished job should

14:46look like. Then, right after you do that, you can add a gate, which is basically a test your agent has to pass before it's allowed to say it's done. So, in a second message, you can type {slash} goal gate add, and then a command that checks the work. In this case, one that checks every single link on that page actually works. Since this isn't an actual link that it's going to create, I'm not going to add that, but that is something I would use if you're working on, say, a website. And the thing is, if that test fails, it gets sent straight back to work with the

15:18exact error. So, now it can't just tell me it's done. It has to actually prove it. Oh, and here's something that I always do for important goals. I send this to Hermes first. I want to, then you state your goal, "Help me write the best {slash} goal prompt to send to you along with a {slash} goal gate add command that I should send along with it, so that you know what a goal is done." And it writes both of them for me. So, all I have to do is copy and send that to Hermes. But with that, when you type a {slash} goal command, it might ask you some questions along the

15:51way. Go ahead and fill those out, and it will only stop the job once it is completely done. And by default, a goal only gets 20 turns. So, it can't run forever and drain your wallet. And while it's working, you don't have to just sit there and watch it. You can use {slash} steer to nudge it in a different direction without stopping it, or {slash} by the way to ask it a side

16:15question without interrupting the job. That handles one big job, but one agent can only carry so much on its own. When you hand Hermes a big job, it can split it up and pass the pieces to helper agents. Those are called sub agents. They run at the same time in the background, up to 10 at once by default,

16:36and each one starts with a clean slate. So, your main chat doesn't get filled up with all their back and forth. Only their final answer comes back to you. Now, here's the money hack, and it's straight from the Hermes docs. On a big job, those helpers burn most of your tokens. So, you want your smart model doing the planning and the helpers running on a cheap one. So, in my Telegram chat, I just sent this to my agent. And that's it. My main model still does the planning, and the helpers do the grunt work for a fraction of the price. And that one setting is how you

17:11run a whole team of agents without paying for a whole team. And if you've got different parts of your life, like your business and your content, give each one its own profile. Each profile gets its own memory, its own skills, and its own settings. Just never point two agents at the same Hermes folder, because they'll keep piling their memories into each other's until neither one is the agent that you set up. I went deep on building a team like this in my bot mode video, so I'll link it up here

17:41for you to watch later if you want to. That lets your agent do a lot more, but the more you add to it, the slower it gets. Every message you send, your agent carries a backpack with it. It's memory, its personality file, its list of skills, and the instructions for every single tool and app you've connected to it. So, the more you connect, the heavier that backpack gets. Every reply costs you more. It gets slower, and the model has a harder time focusing on what you actually asked it. So, whenever my agent starts feeling dumber than it's

18:17used to, this is the checklist I run. And I can do all of it from my Telegram chat. First, I type {slash} status to make sure it's actually on the model I think it is on. Then, I can type context, which shows me how full the chat is and a breakdown of what's taking up space. I can even do {slash} context all, which also gives us a breakdown of what's taking up space all the way down to which skills and tools are costing me the most. And the fix is usually pretty simple. If an app connection gives your agent 30 tools and you only use two of

18:53them, you can tell Hermes to only load those two. And if a chat has just gotten really long, {slash} compress shrinks it down without you having to start over. And I mean, to be fair, Hermes has gotten a lot better at this on its own. It now holds back the instructions for tools you rarely use until it actually needs them. In the October update with version .22, the desktop apps tool instructions went from about 13,000 tokens down to under 7,000. So, that's all seven. If you start a new chat every time you finish a task, keep your expensive model for the work that counts, turn anything you do twice into

19:33a skill, write scheduled jobs that only message you when something's worth your time, put a gate on your big jobs, and let the cheap helpers do the heavy lifting, and keep that backpack light. You're already getting way more out of Hermes than most people do. So, tell me in the comments, which of these seven were you doing? And which one are you going to fix first? I'm genuinely curious. And if you want your agents to actually remember you across every tool you use, try Honcho with my code for $10 in free credits. That's the first link in the description. And as always, every command and prompt from this video is

20:12free in the second link in the description. Open it or just hand the whole thing to your Hermes and let it set this up for you. If you enjoyed this video, then leave a like. Or if you're new to my channel, then subscribe, because I've got a ton more content like this coming your way. Oh, and would you look at that? The algorithm gods seem to think that you're really going to enjoy this next video as well. So, click it,

20:35and I'll see you there.