AI second brain: what the Karpathy wiki and always-on agents mean for your company

What is an AI second brain?

It is a folder of plain text files about your work — who you deal with, what you promised, how you write, what each project needs — that an AI writes and maintains for you, and reads before it does anything.

The idea went mainstream in April 2026, when Andrej Karpathy described how he now spends more of his AI time building knowledge than code. His agent turns raw sources into an interlinked Markdown wiki, keeps it consistent and files the answers to his questions back into it, so the wiki grows with every use. He says he rarely edits it by hand. Millions of people read the post, and dozens of open-source versions appeared within weeks.

The difference from a normal chat is simple. A chat starts from zero every time and you re-explain your work. A second brain starts from everything it has already filed.

Why files, and not the assistant’s memory?

Because memory is a summary and a file is the record. Built-in memory in assistants keeps short notes the model chose to remember. A file keeps exactly what was written, you can open and correct it, and any other agent or tool you add later can read it.

The guides that work split facts in two:

• Slow facts — who people are, how you write, what each project is for — live in files.
• Fast facts — status, dates, amounts — are checked at the source every time, because a copy goes stale.

This is the same rule we use in company brains: the answer is checked against the original message before it is given.

What changed with always-on agents?

Until recently the popular setups lived on a laptop, so the brain stopped when the lid closed. In August 2026 SpaceXAI launched Grok Bot, where AI agents work on a persistent cloud computer with a browser, file system and terminal, and scheduled routines keep running while you are offline. It is now included with SuperGrok, Cursor Pro and Cursor Teams plans.

That turns the second brain from a notebook into a routine: it can file new mail at night and have a summary ready in the morning.

What does a working setup look like?

A widely shared seven-step guide on X shows the shape well. In our words:

1. A profile from what you already sent. The AI reads a batch of your sent emails and writes two files: who you work with and what you owe them, and your writing rules with real examples. Drafts then sound like you.
2. One folder with a fixed layout. Raw material that is never edited, a file per person and per project, an output folder for drafts, an index and a change log.
3. Nightly filing. Anything you label in your inbox is copied in and sorted overnight.
4. A morning page. One page, not an inbox: new facts with links, guesses marked as guesses, what was drafted, what waits for your approval, open questions.
5. Skills. Once a task comes out right by hand — a plan from a note, a reply from a thread — it is saved and reused.
6. A rules file. Every correction becomes one line the AI rereads before each task, so the same mistake does not happen twice.
7. A weekly review. What changed, what is drifting, what you decided and forgot.

The common thread is restraint: the agent drafts and files, and nothing is sent, bought, deleted or published without a yes.

What are the risks?

• Access. The brain reads your mail and calendar. Read what the vendor’s terms say about storage, training and deletion; for Grok Bot, account and data settings follow Cursor’s.
• No walls between agents. On Grok Bot all of your agents share one computer, files and logins, and the documentation says not to treat separate agents as a security boundary.
• Actions. An agent that can send email can send the wrong one. Keep approval on anything that leaves the folder.
• Cost. Frequent routines use your plan quickly; a few runs a day is usually enough.
• Personal data of others. Your folder will hold colleagues’ and clients’ details. In a business context that is still processing under UK GDPR.

When is a personal brain not enough?

When the knowledge belongs to the team, not to one person. A second brain knows what you read and wrote; it does not know what a colleague agreed on a call last week, and it should not show your inbox to everyone who asks it a question.

A company brain applies the same ideas at team level: sources are the shared channels, drive and call transcripts; knowledge is a graph of people, projects and decisions; every answer is checked against the original message; and access mirrors what each person can already see. We compare the options in company brain vs knowledge base vs Copilot and cover data handling in keeping company data safe in an AI knowledge assistant.

How can you start this week?

1. Try it on yourself first. One folder, a profile file from your sent mail, and a rules file. You will learn more in a week than from any article.
2. Write down the ten questions your team asks each other most. If the answers live in shared chats and calls rather than in your own inbox, that is a company-brain problem, not a personal one.
3. Check your data is reachable. Calls recorded and transcribed, work chat in channels, documents in one drive. The free readiness scorecard takes five minutes.