Second Brain: The External Memory That Catches Your Contradictions
Working with AI takes brains. It’s not “dump everything on a robot.” It’s a collaboration of two minds, sometimes more. That’s exactly the joy of it: you talk to someone who is smarter than you in some area, you complement each other, and you make something that’s impossible alone. Before, I only had that with people. Now talking to AI gives me much more, because it has become a second brain for me.
But “second brain” sounds nice until you actually start building one. Then it turns out this is not about magic, but about mechanics: what flows where, who holds it, and what surfaces when everything ends up in one place. That’s what I want to tell you about, from my own experience.
Videos piled up, and time leaked away
I used to watch videos about AI and think: I should remember this, it will come in handy. I saved those videos to Evernote, into different folders and notes. And I forgot. Where exactly it would come in handy, I forgot that too. Videos pile up, the useful stuff settles somewhere in a feeling of “I’ll apply it later.” Later I don’t apply it, because I don’t remember where, and I don’t remember where from.
I have several agents in production, and each one needs constant tuning for new ideas from the industry. Neuroviz, 12 sub-agents. Personal Brand, a master plus sub-agents and skills. AI Mentor, the same setup but for different tasks. It’s impossible to hold in your head which video suits which agent, and which file to write it into.
And here is what I did. Between myself and the working agents I put a consultant: a separate Obsidian vault and a CoWork agent that reads my architecture. Where each agent lives, what it does, which files it runs on.
I give it a training video. It watches, breaks it down, and gives back not “generally useful to study,” but concretely: which agent, which file, what to add. And a ready prompt right away, take it and paste it.
The first video (Karpathy on the LLM Wiki, 17 minutes) turned into 7 recommendations. Three for Neuroviz, three for Personal Brand, one for AI Mentor. Each one a separate file with ready text and an explanation of why here and why now.
AI as executor and AI as advisor, two different levels
Here I caught a difference I hadn’t seen before. AI as a task executor is the first level: do this for me. AI as a consultant on your own AIs is the second. Between you and the working agents a layer appears that catches knowledge from outside and melts it into upgrades of your own system.
The difference in practice is simple. A video doesn’t settle into “later,” it becomes a concrete edit in a concrete file right away.
I put it all in one place, and the contradictions surfaced
Then it got more interesting. I now had three serious agents, each for its own thing: one for the business, one for the personal brand, one a personal mentor. And each knew me by its own slice. The business agent, about role and processes. The brand agent, about how I write and what about. The personal one, about what’s under the hood. Three different pictures of me, and not one whole one.
I brought them together into one second brain. I didn’t merge the agents into one, they still live separately, each with its own thing. I just made one profile that pulls out of each of them what it knows about me, into one place. Reading in one direction: from the agents into the brain, back only when needed.
It sounded like a small thing until I assembled it. It turned out the agents contradict each other on facts: the role was named differently, a couple of facts about the business had drifted apart. While everything sat in separate repositories, the discrepancies weren’t visible. I brought them together, and they surfaced at once.
This, for me, is the main thing. The very act of assembling in one place works as a double check. External memory doesn’t just store, it collides your versions of yourself and shows where they don’t match. In your head these contradictions live calmly for years, each in its own context. On paper, in one file, they have nowhere to hide.
The structure is held not by me
There was one more surprise I didn’t expect. I was always tense that I couldn’t grasp the structure of my own storage: which folder is where, what goes where. And then I caught it: I don’t need to hold that in my head at all.
The structure is held by a separate agent connected to Obsidian. Everything gets dumped in there: summaries from the agents, my own thoughts and ideas, useful videos, transcripts. And it sorts it all out itself. My part is just to say what’s in my head. It does the arranging.
This flips the usual role. Normally a tool demands that you hold its structure in your head: remember where everything is. Here it’s the opposite, the tool holds the structure, and I free up my head for the one thing only I can do, think.
It only works with open files
One important caveat, without it none of this comes together. This only works if the agents are open: plain files you can read and edit. The consultant layer physically goes into your architecture and writes into it. You can’t add anything from outside to a closed box. A Custom GPT, for example, is out: you can’t let another agent read and rewrite its insides.
So this whole second brain rests on the fact that the knowledge sits in a simple, open form. Not in someone’s app, but in your files.
A second brain that remembers for you
A second brain that knows me better than anyone and remembers for me what I forgot. A little creepy, to be honest. But so far I’m delighted.
And here is what’s underneath it. This is not “dumping on a robot” and not a replacement for me, it’s a collaboration of two minds. It holds the structure, remembers the facts, collides my versions of myself and shows where they drifted apart. I think, I feel, I decide what of it is mine.
External memory turned out to be not even about holding less in your head. It’s about seeing yourself whole, in one place, with all the contradictions that hide in different corners of your head. That, unexpectedly, is closer to self-knowledge than to automation.