Latest / AI Fire Daily / #35 Robin: The "Brain in a Box" - How Hermes Agent Turns a Cheap VPS into a Proactive, Memory-Rich AI Personal Assistant
Transcript
- 0:00We have a very passive relationship with AI right
- 0:02now. Think about your daily routine. You open
- 0:05a tab, type a prompt, and wait for an answer.
- 0:08Then you take that answer and do the actual work
- 0:11yourself. But imagine an AI that doesn't just
- 0:14wait around for your questions. Imagine one that
- 0:17actually acts. It runs tasks in the background.
- 0:20It checks your servers. It messages your phone
- 0:22with updates while you sleep. It is a massive
- 0:24behavioral shift, honestly. We are moving from
- 0:26an AI that just answers to an AI that... takes
- 0:30independent action. It's the difference between
- 0:32having a smart encyclopedia and, well, having
- 0:36an actual employee. Okay, let's unpack this.
- 0:38Welcome to the Deep Dive. Today, we were looking
- 0:41at a huge stack of GitHub documentation, developer
- 0:44blogs, and Reddit threads. All about a project
- 0:47called Hermes Agent. Exactly. Our mission is
- 0:49to explore what it takes to build a self -hosted
- 0:52AI personal assistant. we're going to cover what
- 0:54hermes actually is we'll see how it compares
- 0:56to heavyweights in the space too right and we
- 0:59will break down the five core pillars of its
- 1:01brain and crucial for anyone listening we're
- 1:05going to look at how to safely set it up yeah
- 1:07so it works strictly for you without you know
- 1:10breaking your computer exactly so let's start
- 1:12with what hermes actually is because the repositories
- 1:15are going viral right now They really are. I
- 1:18mean, for a long time, we treated AI like a very
- 1:20smart search box. You asked a coding question,
- 1:23you copied the answer, you pasted it into your
- 1:25editor. You were the one moving the data. It
- 1:28was useful, but it was highly manual. You were
- 1:30still the bottleneck. Exactly. Now, looking at
- 1:33the developer community, user expectations are
- 1:35fundamentally shifting. People don't want a search
- 1:38box anymore. They want a personal operator. They
- 1:40want an AI that remembers their preferences.
- 1:43They want it to understand their ongoing projects.
- 1:46And Hermes isn't just another chat bot you log
- 1:48into. It is an open source framework for building
- 1:52your own personal assistant. And self -hosted
- 1:55is a key term here. Yeah. You don't just log
- 1:58into a website. You run it yourself. You can
- 2:00put it on a VPS. That is a rented remote computer
- 2:03running constantly in the cloud. Right. Or you
- 2:05can run it on a Mac Mini sitting on your desk.
- 2:07You could even run it on an Android phone using
- 2:10an app called Termux. Right. And because it lives
- 2:13on your infrastructure, it is always there. It
- 2:15doesn't go to sleep when you close a browser
- 2:17tab. It is a persistent entity. It is also multi
- 2:21-channel. This is the part that fascinates me.
- 2:23You don't open a special AI app to talk to it.
- 2:27You talk to it via Telegram, Discord, Slack,
- 2:30or WhatsApp. The documentation points to Telegram
- 2:33as the easiest starting point. And that choice
- 2:36changes the psychology of how you use it. Yeah,
- 2:38it makes Hermes feel like a real human assistant
- 2:40living inside your phone. You just text it, you
- 2:43are at the grocery store, you get an idea, and
- 2:45you just send a message to your bot. And it actually
- 2:47uses tools. Look, standard AI is like a smart
- 2:50dictionary. It knows a lot of things, but it
- 2:52just sits on the shelf. Right. Hermes is like
- 2:54giving an assistant their own desk. You give
- 2:56them a set of tools, a memory system, a calendar,
- 2:59and a phone number. It can run actual terminal
- 3:01commands. It uses browser automation to look
- 3:04at websites. It even has vision to analyze images
- 3:08you send it. That is the crucial difference here.
- 3:10The workspace is where it runs. The tools are
- 3:14what it uses to manipulate that workspace. The
- 3:17calendar is for its scheduled tasks. It isn't
- 3:20just generating text. It is executing logic.
- 3:23Whoa. Beat. I mean, imagine an assistant that
- 3:26remembers your exact workflows. And it can run
- 3:29them in the background without you typing a thing.
- 3:31It's wild. It's pulling reports and formatting
- 3:34data while you literally go for a walk. That
- 3:36is the real promise here. It isn't about writing
- 3:38a better poem. It creates reusable workflows
- 3:40for the tedious digital chores you repeat often.
- 3:43But this makes me wonder, does this mean we're
- 3:45replacing our standard chat GPT habit entirely?
- 3:48No, it just moves AI from conversation to...
- 3:52actual background task execution okay that makes
- 3:55sense it is a different kind of tool entirely
- 3:56and that brings us to the broader landscape yeah
- 3:58if we're going to invest time setting this up
- 4:01we need to know where it fits we don't want you
- 4:03using a hammer to drive a screw exactly let's
- 4:07look at how armies compares to other big names
- 4:09in the agent space specifically claude code and
- 4:12open claw right if we connect this to the bigger
- 4:15picture the ai agent space is fragmenting really
- 4:18fast Six months ago, an agent was just a buzzword.
- 4:22Now, different tools have very specialized jobs.
- 4:26Let's start with Cloud Code. Going through the
- 4:28developer blogs, it is clear this is built for
- 4:31focused, deep work. desk work. Exactly. Cloud
- 4:34Code is what you use when you are sitting at
- 4:36your computer staring at a massive code base.
- 4:39It excels at local environment management. It
- 4:41edits your files, manages your repositories,
- 4:44and works deep inside your terminal. It is a
- 4:47dedicated coding partner. But it is not a mobile
- 4:49first assistant. You aren't going to text Cloud
- 4:52Code on WhatsApp while you're driving to remind
- 4:54you about a meeting. No. Or to summarize a news
- 4:57article. Right. It is heavy machinery. Then on
- 4:59the other end of the spectrum, you have something
- 5:01like OpenClaw. The repositories show this is
- 5:03a much heavier system. Okay. It is basically
- 5:05a multi -agent gateway. Wait. A gateway system.
- 5:09So it connects messaging apps, multiple specialized
- 5:12agents, and external plugins into one massive
- 5:15architecture. Yeah. That sounds incredibly complicated.
- 5:18It is. It is powerful, but it can feel very brittle.
- 5:22It is really designed for power users, startup
- 5:24founders, or developers building complex commercial
- 5:27platforms. So when you have that many moving
- 5:29parts, if one plugin updates, the whole house
- 5:33of cards might fall. Exactly. It requires constant
- 5:36maintenance. Which leaves us with Hermes Agent.
- 5:38It sits in a very specific practical lane. It
- 5:41is lightweight. It is telegram first. And it
- 5:44is designed to be a self -improving personal
- 5:46assistant. Hermes is for when you are walking
- 5:49around with your phone, but you still want your
- 5:51digital work to move forward. You want something
- 5:53that grows with your specific habits. And as
- 5:55a bonus, the architecture is great for testing
- 5:57new open source language models. Yeah, definitely.
- 6:00But hold on. If I am using cloud code for my
- 6:03programming and I am setting up Hermes for my
- 6:05daily schedule, aren't I just giving myself a
- 6:07management headache? I mean, it feels that way
- 6:10at first, but separation of concerns is actually
- 6:12it. s practice here right you wouldn't ask your
- 6:15accountant to also design your logo exactly you
- 6:17want specialized tools for complex tasks hermes
- 6:20acts as the project manager while claude is the
- 6:23factory worker so if i'm just coding an app is
- 6:26hermes the wrong tool yes use claude code for
- 6:29deep coding and hermes for daily assistant tasks
- 6:32got it now we really need to look under the hood
- 6:35to understand how hermes manages to be so fiercely
- 6:39independent we have to examine its brain Yeah,
- 6:42the architectural docs break this down into five
- 6:44core pillars. These five pillars are the exact
- 6:47mechanisms that turn it from a reactive, forgetful
- 6:50chatbot into a proactive, reliable agent. Pillar
- 6:53number one is memory. And honestly, this is the
- 6:56biggest friction point with AI right now. Most
- 6:59AI models wake up with total amnesia. Oh, absolutely.
- 7:02Every single section starts totally fresh. That
- 7:04is incredibly frustrating. You don't want to
- 7:06explain your business model, your project context,
- 7:09and your personal communication style every single
- 7:12time you open a chat. Ermi solves this at the
- 7:14root level using two specific markdown files.
- 7:17Let's break those down. The first is the user
- 7:19file. This file is entirely about who you are.
- 7:23Right. It holds your professional role, your
- 7:25preferences, and your writing style. You can
- 7:28literally write in the file, keep my email simple,
- 7:31warm, and direct. And the mechanism here is crucial.
- 7:34Every time Hermes wakes up to do a task, it secretly
- 7:38reads that file first. It gives itself a crash
- 7:40course on who you are before it even sees your
- 7:42prompt. The second file is the memory file. This
- 7:45holds your project context. Yeah. So business
- 7:47information, server environment details, recurring
- 7:49facts, stuff that doesn't change daily. Yeah.
- 7:52But the docs warn you have to be careful here.
- 7:54You really do. You only save durable facts, not
- 7:57temporary task statuses. If you save every little
- 8:00to -do item, the file gets bloated and the AI
- 8:03gets confused. Exactly. Pillar two is skills.
- 8:05If memory dictates what the AI needs to remember,
- 8:08skills dictate how it repeats a task. They act
- 8:11as reusable recipes for your repeated digital
- 8:13chores. These live in individual skill files.
- 8:15They contain step -by -step instructions. They
- 8:17also use YAML front matter. Let's define that.
- 8:20Hidden code tags that tell the AI when to trigger
- 8:23skills. So instead of just guessing how you want
- 8:26to report formatted, it reads the hidden tags,
- 8:29sees the trigger word, and pulls up the exact
- 8:31blueprint you created last week. And I have to
- 8:33admit, I still wrestle with prompt drift myself.
- 8:37Yeah. Beat. Yeah, it's so common. You know, having
- 8:39to constantly rewrite instructions because the
- 8:42AI slowly forgets what I want it to do by the
- 8:4510th message in a chat. This skill system. completely
- 8:48fixes that. It absolutely does. You stop asking
- 8:51the agent to improvise every time. You give it
- 8:53a rigid track to run on. Which brings us to pillar
- 8:55three, the soul. This lives in the soul file.
- 8:59This one's fascinating. This file gives the agent
- 9:01its actual personality and tone. You can make
- 9:04it warm and empathetic. You can make it highly
- 9:06sarcastic. Or you can make it strictly technical,
- 9:09returning only code and no apologies. And this
- 9:12matters deeply if Hermes is, for example, drafting
- 9:15customer support replies for your business. A
- 9:18generic robotic assistant sounds totally forgettable.
- 9:21You want it to sound like your brand. How does
- 9:23the sole pillar actually change the output practically?
- 9:26It dictates the tone, preventing your assistant
- 9:28from sounding like a generic corporate robot.
- 9:31Right. It literally changes the weights of the
- 9:33words the model chooses. Pillar four is cron
- 9:36jobs. Let's define that jargon too. Scheduled
- 9:39tasks that run automatically at a specific time.
- 9:42This is the feature that turns Hermes from a
- 9:44reactive tool into a proactive operator. You
- 9:48aren't just messaging it, it is messaging you.
- 9:50You schedule a market research summary for 7
- 9:53a .m. or database backup every Friday at midnight.
- 9:56It just happens without a prompt. And pillar
- 9:59five is the self -improving loop. This is the
- 10:01absolute magic of the system. It is the process
- 10:04of turning your manual corrections into permanent
- 10:07skills. You ask Hermes to do a piece of work.
- 10:10It does it, but maybe the output is messy. You
- 10:12correct it. Then Hermes saves that newly improved
- 10:16workflow. It searches its past sessions to keep
- 10:19continuity. It literally rewrites its own instructions.
- 10:21So the next time it runs that task, the output
- 10:24is simply better. It learns from its mistakes.
- 10:27Okay, we are going to talk about how to safely
- 10:29set this architecture up next, right after this.
- 10:33Sponsor, here's where it gets really interesting.
- 10:36We know how its brain works with the markdown
- 10:39files and the cron jobs. Now, how do we safely
- 10:42give it a body in the real world? Because we
- 10:44definitely do not want an autonomous AI accidentally
- 10:48destroying our personal files. Setup and security
- 10:50are absolutely critical here. Remember, Hermes
- 10:53can run terminal commands. That means it can
- 10:56delete files. It can install software. That is
- 10:58incredibly powerful, but very risky. The documentation
- 11:01strongly recommends running it on a separate
- 11:04VPS or an isolated Docker container. Let's define
- 11:07Docker container, a secure digital box keeping
- 11:09software isolated from your main computer. Think
- 11:11of Docker like putting the AI in a soundproof
- 11:14room inside a shared house. If the AI makes a
- 11:17massive mess, the house stays perfectly clean.
- 11:19You just delete the room and start over. It keeps
- 11:21everything safely organized. You log into your
- 11:23remote server, install Docker, create a clean
- 11:25folder, and then run the official installation
- 11:27script. Right. You do not just download it to
- 11:30your desktop. During that setup, you connect
- 11:32it to Telegram using a tool called Botfather.
- 11:35This gives you a unique bot token. But the community
- 11:38forums highlight a massive security warning here.
- 11:41They do. You absolutely must lock the bot strictly
- 11:45to your specific Telegram user ID. Yes. If you
- 11:49skip that step, anyone who finds your bot on
- 11:51Telegram can start texting it. And since it has
- 11:54terminal access, random strangers could be running
- 11:56malicious commands on your private server. It
- 11:59is a huge vulnerability. Another strict rule.
- 12:01Never, ever paste your API keys, your passwords,
- 12:05or your tokens directly into the chat window.
- 12:07You store those secrets safely in .env files.
- 12:10That's a hidden digital safe used specifically
- 12:13for storing your private passwords. And you use
- 12:15a .gitignore file to protect them from being
- 12:17uploaded to public code repositories. Clean security
- 12:20habits from day one are essential. The developers
- 12:23also recommend disabling any tools you don't
- 12:25actually need yet. If you aren't using browser
- 12:27automation, turn it off. Limit the attack service.
- 12:31So why shouldn't I just run this locally on my
- 12:32MacBook right now? An autonomous agent with terminal
- 12:35access could accidentally break your personal
- 12:38computer files. That is a terrifying thought.
- 12:40So we put it in a safe space, we lock it down,
- 12:43then we build our first workflows. Yeah, and
- 12:46the community guides say you should not build
- 12:48complex multi -step automations on day one. Start
- 12:52very small. The very first workflow you build
- 12:55should be connecting a private GitHub repository
- 12:58for daily cron job backups. This makes total
- 13:01sense. You are backing up the memory and skill
- 13:03files. You are literally saving the brain and
- 13:06habits of your assistant. If the server crashes,
- 13:09you don't lose the weeks of training you put
- 13:11in. The second recommended workflow is a 7 a
- 13:13.m. morning briefing. You set a cron job for
- 13:16Hermes to pull your top priorities, scan relevant
- 13:18industry news, and flag urgent calendar reminders.
- 13:22It sends it to your phone right when you wake
- 13:23up. It makes Hermes feel genuinely helpful immediately.
- 13:26You get a clean, customized start to your day.
- 13:29The third recommended workflow is a weekly server
- 13:31health check. Reviewing disk usage, checking
- 13:34running services, and scanning error logs. Hermes
- 13:36lives on a server, so ask it to manage that server.
- 13:39Exactly. Checking disk usage isn't just asking,
- 13:42is the disk full? It's having the AI run the
- 13:46command, read the output, notice that a log file
- 13:49is eating up 80 % of the drive, and texting you
- 13:52to ask if it should clear it. It built a great
- 13:54maintenance habit early on. We have the setup,
- 13:56we have the safety checks in place. What does
- 13:59this actually look like in the wild when it starts
- 14:01working along? This is where we look at the real
- 14:05-life applications. Sure, you can use it as a
- 14:07simple personal assistant for reminders, but
- 14:10the architecture scales up beautifully. You can
- 14:13use it as a full business operator. monitoring
- 14:15a Discord community, summarizing hundreds of
- 14:18YouTube comments to find common questions, or
- 14:20drafting technical support replies based on your
- 14:23actual documentation. What's fascinating here
- 14:25is that the value isn't day one perfection. You
- 14:28have to adjust your expectations. The real value
- 14:30is day 30 automation. Exactly. Think about a
- 14:33content creator workflow. Hermes could draft
- 14:35scripts, pull research, and create data diagrams.
- 14:38You can give it a cron job and say, every Friday
- 14:41at 4 p .m., review my audience comments from
- 14:44the week and create 10 new content ideas. It
- 14:47just happens. And you can interact using voice,
- 14:49which is huge for accessibility and speed. You
- 14:51can be out on a walk, not looking at a screen,
- 14:53and just brain dump a chaotic voice message into
- 14:55Telegram. Hermes will transcribe it, parse out
- 14:58the actionable items, and create a structured
- 15:00task list in your system. It also has vision
- 15:02capabilities. You can feed it UI screenshots
- 15:05from an app you were building or complex data
- 15:08diagrams and ask it to check for visual errors.
- 15:11Or you can use it as a dedicated, relentless
- 15:14research assistant. You ask it to monitor specific
- 15:16AI news feeds during the week. Compare three
- 15:20different software tools. Pull the pricing. format
- 15:23it into a table, and save the findings directly
- 15:25to its memory. And this all feeds right back
- 15:27into that magic self -improving loop we talked
- 15:29about earlier. You start with one real messy
- 15:32task. You let Hermes attempt to help. You review
- 15:35the output and you correct it. You say, no, don't
- 15:38use bullet points. Use numbered lists and keep
- 15:40the summaries under two sentences. And then you
- 15:43save that exact pattern as a permanent skill.
- 15:45What happens if the first research reported generates
- 15:47is completely useless? You correct it, then tell
- 15:50it to save that improved format as a skill. It
- 15:53is so incredibly practical. You are not just
- 15:56sitting around waiting for the big tech companies
- 15:58to make the base AI model smarter. You are actively
- 16:01teaching your specific instance how to work perfectly
- 16:06for you. Beat. you build a fundamentally better
- 16:09way of working together. By offloading these
- 16:11routines, it drastically reduces the number of
- 16:14small, exhausting decisions you have to make
- 16:16every single day. So let's bring this all together.
- 16:19Hermes Agent and tools like it represent a fundamental
- 16:22shift in computing. We are moving from renting
- 16:24intelligence one prompt at a time to owning an
- 16:27evolving operator to sex silence. It builds its
- 16:31own memory. It builds daily habits. It builds
- 16:33complex workflows tailored exclusively to the
- 16:35messy reality of your life. It is definitely
- 16:38not magic out of the box. I think that's important
- 16:40to stress. It requires genuine guidance and discipline.
- 16:43You have to actively prune its memory and take
- 16:46the time to create those foundational skills.
- 16:48But the compounding return on your time is immense.
- 16:52Once you get those skills dialed in, it operates
- 16:54silently in the background, constantly pushing
- 16:57your work forward while you focus on the big
- 16:59picture. It essentially becomes a custom, lightweight
- 17:03operating system for your professional life.
- 17:06Before we wrap up today, I want you to think
- 17:08about this. Look closely at your daily routine
- 17:10right now. What is the one tedious digital chore
- 17:14you do every single morning? Pulling data, checking
- 17:17a dashboard, formatting an email that you can
- 17:20hand off to an autonomous assistant right now.
- 17:22Thank you for joining us on this deep dive. We
- 17:24covered a lot of technical ground today, but
- 17:26hopefully it gave you a blueprint for the future.
- 17:28So what does this all mean? It means your new
- 17:31assistant is waiting for instructions. Out hero
- 17:33music.