How should I set up an AI second brain for my business?

I've had so many stops and starts with a second brain. I built a private one five times. Then I looked at what was in it and realized 98% was just company stuff.

"I'm simplifying everything to say everything is either business or it's personal. My personal stuff is completely separate, a hundred percent. Everything else... is one thing and everyone has access to all of it."

What is a company brain, actually?

Files. Markdown files, which are basically Word documents. The agents read those documents to understand everything about your business. Nothing fancier than that.

The point is that the brain knows everything you're doing and talking about, and it's independent of Claude or Codex. Plenty of businesses worry about vendor lock-in, and they're right to. If the knowledge lives in your own files, you can switch tools without losing it. And nobody has to change behavior. You still use Claude or Codex, hopefully Slack, and the brain fills itself in the background.

Should I have a personal brain and a team brain?

Yes, but draw the line at business versus personal, not at "mine" versus "the team's." Everything about the business goes in one shared brain that every teammate and agent can read. Personal material stays fully separate. Don't make the two writable to each other.

When I checked, about 95% of my second brain and my business email had no sensitivity at all, so one business chief-of-staff agent can read it and the team can share it. Truly personal things get a separate agent on my personal laptop.

For a team, I think of it as two layers. The company brain everybody reads, but only the owner changes. Then each person gets an individual brain the system writes to automatically from what they're doing.

The rule underneath it: you can't have two sources of truth. The moment the same fact lives in two places, one of them goes stale.

How does it stay up to date without me?

• Your AI chats. Set up properly, a daily job pulls what's valuable from your Claude and Codex chats and adds it to the brain.
• Your calls. I record all my Zoom calls with Fathom, which is free and auto-joins spur-of-the-moment calls that other notetakers skip. The transcripts get synthesized into the brain. When I asked my agent for five content ideas about local newsletters, it came back with ideas from things I'd said on past calls.
• Other smart people. You can pull in content that isn't yours, like a marketing expert's daily podcast, synthesize it and keep only what relates to your business. Now the brain gets smarter from what you do and from what you think smart people are doing.
• A cheap model for the backlog. Backfilling old transcripts doesn't need the top model. I run a cheaper one at 20 to 50 a day until it's done.

How do you keep agents from drowning in it?

Curate what they see. Everything can live on disk, synced to GitHub so it's permanent. But I'd only point a curated Obsidian vault at the valuable stuff: what the product is, the best marketing strategy, what we learned in the last six months. Logs and things your AI creates that you need to keep but your brain doesn't need to know stay out. Then tell your agents they can only reference the vault.

My promotion rule: an agent scores each new item, like a call transcript, against a document I wrote on what makes something relevant. An 8 out of 10 or higher gets promoted to the vault. A 7 or lower stays in the files. It'll make mistakes, but it's not on me as a human to sort it. Putting everything everywhere is the real mistake. An agent with too much information shuts down, and that's worse than nothing.

Why did my earlier second brains fail?

An earlier version had plumbing that worked fine, but the knowledge itself was about 2 out of 10 useful. It was mostly navigation and technical receipts, not substance. A brain full of logs about the brain isn't knowledge.

Even the next rebuild, on the smartest model available, was still a disaster. It had about 12,000 pages, which sounded wrong to me by a factor of at least 10. When that happens, don't patch. I asked for a post-mortem and told the AI to ignore the current design and start from the best practice for a company second brain. The value is highest when every agent is connected to it and uses it on purpose.

My rebuild keeps raw evidence (transcripts, notes, source files) separate from curated knowledge (the actual answers and decisions). They live in separate repositories.

What tools have I used?

• Obsidian. In January 2026 I built an intelligence dashboard in 20 or 30 minutes with Claude Code. It pulled YouTube channels I like, synthesized what mattered for local media and newsletters, drafted LinkedIn and Twitter posts, and pulled in my meeting transcripts. Obsidian is like Notion for markdown files, which is what AI reads, and it can simply be the display layer for a folder on your hard drive.
• Notion. As of October 2026, every podcast episode gets a page with takeaways and social posts. Each client portal pulls metrics from greetform and Meta ads. My email and Fathom call transcripts flow in automatically, so I don't even need to make a note in my CRM. I drop interesting links and they get analyzed with next steps. A checkbox can trigger an agent to act.

One design note from the Notion rebuild: tell the model to avoid walls of text and use visuals. I tried the Notion build with Codex first and it was bad at it. Claude's top model did it well, far better than a smaller model.

One limit: the second brain shouldn't become your production database. For my local directories, Notion is fine as an editing layer, but the production database stays the source of truth. Notion shouldn't become the live database for thousands of listings. Production publishes an editable view, people propose changes, and one controlled writer applies the approved edits.

What do I do when the AI system isn't ready?

Don't let team memory wait on automation. When my agents failed one day in September 2026, I created a plain shared channel called reference and told the team: if you're doing any real work, document it. Shared references, client history in the CRM and agent work stay separate, so nothing depends on the agents being up.

Start small: have the AI interview you and save the result as the first page of your brain.