Barnabas is out of the Cryo chamber. I released him.
Same personality.
But he is no longer pestering me with worthless ideas and catchphrases that make no sense.
That was the problem with the first Barnabas.
He was memorable, but he was not useful… at all. He could brainstorm. But that was about it. And he didn’t even do that well. I still had to remember every decision, move every task forward, write every post, build every feature, and manage every request myself. Barnabas was essentially a bad comedian.
He was a very entertaining reminder that I was wasting my time and money.
Transformation from a goofy capybara to a competent CEO (Chat GPT)
So I got rid of him.
But now he is back, he is better than ever, and he has hired a full team...
Barnabas 2.0 has superpowers, for multiple reasons.
First… he lives on real infrastructure.
He is hosted on a Hetzner German VPS, running in a Docker container, and operating through Hermes Agent. That matters because Hermes is not just another chatbot wrapper. Hermes is an agent that gives him tools, memory, files, scheduled cron jobs, messaging, web access, image generation, video generation, and the ability to work across systems instead of being trapped in one conversation.
Barnabas 2.0 had one of his employees make the chart below....The old Barnabas could never.
The new system. Many agents, one brain.
ChatGPT can answer me.
Hermes can do actual work. And it can do work from anywhere, without being asked. 24/7!
(Hermes is an open source AI “agent” from Nous Research that you can download. It can access “models” like Open AI's Chat GPT 5.5 and Anthropic's Opus 4.8. Barnabas is just personality text file with maybe 30 lines that was created by another open source agent over a month ago. Go back 5 articles if you want more info.)
Anyways, Hermes can read files, write drafts, search old sessions, run scripts, manage cron jobs, hand work to other agents, and come to me when something needs approval. My conversations with it can live in Telegram and Discord, so I can talk to Barnabas from my phone. It can run on the server, connect to my projects, my website, Gmail, Calendar, local files, research tools, coding tools, and use any model you can think of (from simple to state of the art).
Second, I am building him a real brain. A big one.
Not “memory” in the vague chatbot sense where it remembers a few preferences and forgets 90% of what you needed it to remember.
I am building a shared brain that combines the parts I like from Obsidian combined with Andrej Karpathy's LLM Wiki, and Garry Tan’s G-Brain.
Obsidian gives me the part I can actually touch.
It is not magic memory. Obsidian is just a file/note organizer. It's bunch of Markdown (text) files, folders, links, and notes I can open, edit, move, delete, or if Barnabas starts storing weird notes. It's organized and visible. And that matters. If AI is going to use my work as memory, I want the memory somewhere I can inspect it.
The LLM Wiki idea is the next piece. It's just a method used to store the files in Obsidian. Karpathy and others have been pointing toward a simple but important idea... models should not just rummage through a junk drawer of notes. They need organized context. A wiki gives the system a more structured map of projects, people, concepts, decisions, sources, and open questions.
Karpathy's process is actually very simple:
Human and AI saves raw notes and files into an input folder.
Let an AI model organize and categorize them. Then it takes key concepts from the files and creates "wikis" (summary files).
Models can easily search through the summary files based only simple instructions, and only go deeper into raw files if needed.
Efficient and simple.
G-Brain is good for operating memory. Not easy to see or edit but stores and organizes well automatically. Great for: chats, decisions, links, artifacts, and the history of what actually happened.
The power is putting those pieces together.
Obsidian keeps the brain editable. The LLM Wiki structure makes the knowledge easier for agents to navigate. G-Brain is the broader memory layer across my work. Hermes sits on top and lets Barnabas use that memory while he is doing actual work.
Obsidian + LLM Wiki. Interconnected brain with efficient storage and retrieval. (Evantravers.com)
No magic orb. No “the AI just knows.”
Just files, structure, retrieval, and enough plumbing to keep the capybara from making his goofy little guesses.
Third, this brain is not just for Barnabas.
I want Hermes, Codex, Claude Code, Claude chat, ChatGPT, and the rest of my models to stop acting like separate interns who each forget what the other one learned.
If Codex figures something out in a repo, that should not vanish into a coding session. If Claude Code helps design a workflow, that context should be available later. If ChatGPT helps me think through an article, the useful parts should not be buried in chat history. If Barnabas makes a website decision or has one of his agents do research, I should be able to find all of it easily and quickly.
One shared brain changes that.
Barnabas’s new and improved brain.
It means I can send Barnabas a task from my phone, pick it up from my laptop, let an agent work on the VPS, and still have the context live in one place. The model changes. The interface changes. The memory does not disappear.
And Barnabas is now the CEO of Dustin LLC.
He has four employees.
Chat GPT
Scout is the purple fox. Scout does the research: AI tools, market shifts, Fanbase Weather opportunities, competitor ideas, sports chaos, and anything that needs evidence before I write or build around it.
Scribe is red. Scribe helps to write: newsletters, LinkedIn posts, scripts, captions, product updates, documentation, and drafts that need to sound like me. (Scribe still sounds nothing like me. A lot of improvement to be made here.)
Reach is pink. Reach handles marketing and social media: post strategy, account planning, launch angles, audience hooks, promotion ideas, and getting useful things in front of people.
Dev is blue. Dev builds: code, integrations, dashboards, Fanbase Weather improvements, model tests, scoring workflows, and the technical pieces that make the system less theoretical.
Barnabas does not just talk about them.
He built a Mission Control to monitor them.
Mission Control Homepage
That is the dashboard where I can see the agents, their work, scheduled tasks, content, blockers, and what needs me. I can send Barnabas a task, and he can decide where it goes. Research to Scout. Drafting to Scribe. Marketing to Reach. Build work to Dev.
He is already managing more than five social media accounts. He is already helping improve models on fanbaseweather.com. His agents are researching, writing, planning marketing, checking workflows, and feeding status back into Mission Control.
Barnabas is still allowed to be weird.
The personality makes the system easier to talk about. But this time the personality is sitting on top of infrastructure: Hermes, the VPS, Docker, shared memory, Obsidian, LLM Wiki, G Brain, Mission Control, and a team of agents with actual jobs.
Barnabas came out of cryo with the same personality.
But now he has tools.
Now he has memory.
Now he has employees.
And he is finally capable enough to get some impressive stuff done.
🙏 Thanks for reading
🗞️Newsletter: LinkedIn Newsletter
🌎 Site: dustincoledata.com
- Dustin