The hype that is AI in your second brain

I find it very amusing how a new hype keeps cropping up in tech. Right now, in my AI bubble, it’s about the idea of letting an AI model loose in your second brain. Everyone jumps on it right away until the next one shows up, and then we all happily move on. And yes, I’m part of that crowd too. I jump along just as cheerfully, and I’ll probably be up front again for the next hype. Still, I think there’s genuinely something to gain in this one. Either way, my work has become a good deal more productive because of it.

I’ve been writing everything in Obsidian for years now. My notes, my tasks, my calendar, it’s all in there. That’s my second brain in the literal sense. When I want to know something I once wrote down, I search there and not in my head. It grew slowly, and by now I trust it the way I trust my memory, only better, because it forgets nothing. There is one catch: you have to log everything yourself. It took a while to become a habit, but it’s part of my routine now.

A few weeks ago I gave AI access to that brain. Not as a chatbot I ask something now and then, but as something that permanently keeps up with what happens in my vault. Andrej Karpathy described an idea for this that he calls an LLM Wiki. Instead of a model that rummages through your documents again on every question, the standard RAG approach, the AI keeps its own wiki. A set of markdown files that sits between your raw sources and yourself, maintained by the model itself. As Karpathy puts it: the annoying part of keeping a knowledge base isn’t the reading or the thinking, it’s the bookkeeping. And bookkeeping happens to be exactly the kind of thing a model never gets tired of.

That lined up exactly with what I was looking for. But I didn’t want the AI to just get at everything. That’s where my approach differs.

I’ve split up my vault. There are folders the AI may read but never write to. There are folders the AI can’t enter at all, no reading and no writing. That’s where my own thoughts and notes live. And then there’s one folder that belongs entirely to the AI. In there it can create its own folders and files: agents, skills, context, the llm-wiki, config, everything it needs to serve me well. In that one folder the AI builds its own Obsidian graph, inside my graph.

I only specify the outcome. I say what needs to happen and where it should go, but how it gets there as efficiently as possible I leave to the model. That’s the same division Karpathy describes. I own the what, the AI owns the how.

One rule I’ve baked in tightly. Every note the AI writes gets a property in the frontmatter with the model that wrote it, for example model: claude-sonnet-4-6. That way I can see at a glance which notes have been touched by AI, and by which model. I can then filter on that, which helps me judge the content on its merits. It’s a hard requirement for the model, and it works perfectly.

This diagram was generated with AI.

That wall between my side and the AI’s side is the most important part for me. There are two reasons behind it. The first is quality. My vault is full of half-baked ideas, loose fragments, things I write down before I know whether they’re going anywhere. If I throw all of that into the model’s context, I get output that’s just as messy as my raw thoughts. By letting the AI near only the finished part, what comes out stays sharp.

The second reason is more a matter of principle. My thoughts are mine. The thinking itself I don’t want to outsource, and certainly not see polluted by a model that writes in it uninvited. The AI is a tool, but not a co-owner of how I think. The wall keeps that line hard.

That’s also why I now run a local model. I started with Claude Code, then switched to Codex, but with a model on my own machine everything stays with me. That way the wall doesn’t have to lean on someone else’s promise. Keeping a model like that running does ask more of me, precisely because everything moves so fast right now, but it’s worth it to me.

I effortlessly find things that would otherwise have been harder to track down. I have the model write reports on my projects. Within a few seconds I have an overview of where something stands, what’s still open, and how the pieces fit together. That makes my work more productive, the word that comes with every new tech hype. Only, the way I see it, that’s not down to the AI, because everyone has that. It’s down to the context it has, to the fact that I can deploy AI exactly the way I want, with sharp boundaries for where the model may and may not go.

The wall isn’t an original idea, but it’s how I do it. And when the next hype comes along and everyone moves on again, everything I’ve built around it simply stays mine. And the funny part is that with this piece I’m meanwhile contributing to it myself, to the hype I found so amusing at the start.