Ever heard that Obsidian is a second brain? I heard it from other people plenty. I said it myself. I taught it.
Well. Not anymore.
What we’re building today isn’t a second brain. It’s a shared one — a Shared Brain.
Here’s why, and how.
This one leans philosophical. But it’s the principles and the patterns that will let you build your own system.
A Second Memory
Say “Obsidian” and the first thing that comes to mind is “knowledge base.” Great concept: gather all my knowledge in one place so that… so that what?
Any note-taking system is, at best, just memory. More reliable than mine in some ways. More useless in others.
For a zettelkasten — or any system like it — to actually work and pay off, the system alone isn’t enough. You need discipline and processes. Usually those belong to researchers, students, scientists. To people who write books and papers. To people for whom deep work is a natural state.
I tried. I turned Obsidian into a daily routine. Which, by the way, was hard and didn’t happen on the first try. And it still wasn’t a second brain.
Because a brain isn’t memory.
A brain is the set of processes tied to perceiving and handling information. Dozens of them run in parallel in there, and we never even notice.
Yes, memory is an important part of the brain. But only a part.
Knowledge on its own is useless.
First, it goes stale.
Second, on its own it never turns into action.
A library is dead data, even if every book sits in its right place in a beautiful binding.
What I Changed
My Obsidian took on a new meaning when I let AI agents loose inside it. Claude Code first. Then Codex and Gemini. Then OpenClaw and Hermes.
The agents read my notes.
We started working together. We set up processes.
Overnight they summarize my calls, write work logs for each project, hunt for non-obvious connections. Research, digests, news — all of it lands in the Shared Brain.
Here’s what I typically ask:
- find everything in the vault on this project
- remind me what Robert and I agreed on last week
- find the latest draft of the post about X and add this idea to it
- find the summary of the last meeting and make a YouTube thumbnail for the video
- run research on this topic (the result always lands back in the vault)
- build a card for every member of group Y
The brain didn’t become a second one. It became a shared one.
And this isn’t a property of one particular agent on one particular LLM.
Tomorrow I can swap Codex for Grok — and in 10 minutes the agent finds its way around my work and my life.
Internet goes down? Not a problem, I’ve got local models.
What a Shared Brain Is
A Shared Brain is a system of shared context for a human and their AI agents. One memory, open to both you and the AI. It all lives in your files, on your machine. No cloud storage, no expensive subscriptions.
It’s not a knowledge base with agents bolted onto it.
If you just add agents to your knowledge base, what you’ll most likely get is an information graveyard growing out of control. A beautiful graveyard: crypts, obelisks, monuments, and a couple of zombies if you’re lucky.
Karpathy’s Personal Wiki lands in the same plot, by the way — it just adds a lot of new, very handsome headstones. Cross-linked ones.
The difference is the goal.
A Shared Brain exists to work, not to store and accumulate.
Three Layers of Access
Okay, there’s a question.
If it’s my personal knowledge base, but an agent can write into it — is that knowledge still mine?
Can you even call it knowledge?
Yes, with a caveat. Call it the Seneca caveat.
A podcast summary an agent wrote is just a podcast summary. Research I ran through Exa AI or Parallel AI is not my research. A text I didn’t write is not my text. An agent can parse the internet forever, write its own texts, even learn from them.
But it’ll be the one learning. Not me.
Then again — your own head is full of knowledge you never studied. You don’t know how, exactly, your brain runs your internal organs. You don’t know what it’s busy with while you sleep. And that’s fine — that’s how it should be.
There’s room in your head for all of it. Conscious and unconscious knowledge coexist just fine in there. Explicit processes and implicit ones. Maybe a couple of Billy Milligans too.
That’s my Obsidian now. Three layers:
- My knowledge and my texts. This text, for instance. The agent gets strictly read-only.
- The shared layer. Summaries of podcasts I listened to and read, notes, meeting transcripts, people cards. Everybody writes here — me and the agents.
- The agents’ data. Their work logs, overnight research, statistics. That one is read-only to me.
A bit like the common space in a house shared by several people. There are shared resources, and there’s the stuff that’s specifically mine.
So the formula goes: brain = me + my processes + shared memory + agents and their processes.
How This Differs from PKMS
In case you’re wondering, PKMS = Personal Knowledge Management System.
For the system to work, it needs more than “knowledge” in the classic sense of the word. It needs everything:
- strategy and plans
- decisions and what came of them
- data about mistakes
- statistics and work logs from the team
- data about clients, partners, and other people
- meeting transcripts
- raw ideas and scribbles
None of that fits the textbook definition of “knowledge.”
But all of it matters. All of it is what lets you grow and act.
A word on raw ideas.
You lose them the moment you treat your Shared Brain as a knowledge base. Ideas arrive raw, and that’s normal. Waiting for an idea to turn into a proper note will only slow you down.
Let’s move from Personal Knowledge Management to Personal Execution Management. From managing what you know to managing what it’s all for.
A few of my own processes, as examples:
- building individual learning materials for course students
- working out new features against the full body of data on my projects plus user feedback
- writing proposals with meeting transcripts and the history of previous client relationships in context
- distilling ideas out of the way something got phrased on a call
- prepping for live streams on a week’s worth of accumulated context
Can you copy these processes?
You can… to what end?
Everyone works differently. Copying each other’s processes is pointless. But you can and should take inspiration from the approaches and the patterns.
How Do You Start?
1. Gather your context in one place that’s fully yours. The files live with you. Obsidian, Logseq, a plain folder of files — the format matters far less. What matters is that it’s yours: swap the agents or the tool itself, and the data stays.
2. Agree on the rules with your agents. What they read, what they write, where they never go at all. You can set this up together with any agent. In conversation. But the final call is yours. No rules, no system — you get chaos.
3. Set up your first process that runs without you. Call summaries, an overnight project log, a retro analysis, a link-hunter — anything that turns context into action. A base with no process is a library. And a library is dead data.
The Disclaimer
Nothing is easier than asking agents to make you a mountain of content. But you always want to know the answer to “why?”. Don’t let the brain grow unchecked. Yes, it’s a Shared Brain, but it’s your Shared Brain.
And one more thing.
I’m hand-building what will probably work out of the box someday. All of it will ship as apps and plugins.
Is it worth figuring out now?
My answer is yes. Not to be early, but to understand.
When it does work out of the box, you’ll already know what’s inside, how it works, and where it breaks.
“You can delegate thinking — but you can not delegate understanding.”
By the way, a Shared Brain answers where your context lives, and how.
Compound Context answers why it’s worth accumulating.
Go read that one.
Oh, and — go check: there may be an open enrollment right now for the course where we build these systems together and learn to work with AI.
The course is here: https://anti-magic.com