Ask anyone in your company where the latest pricing is, why you dropped that supplier in 2024, or how a customer in Finland gets onboarded. Someone knows. The problem is that it is always someone, and the someone changes.
"Second brain" started as a personal productivity idea: a trusted system outside your head where you capture, organise and reuse what you learn. The company version takes the same idea and asks a harder question. What would it mean for the business itself to have a memory that does not walk out of the door at five o'clock, or for good in the spring?
What it actually contains
Not everything. A second brain that contains everything is a data lake, and nobody can answer a question from a lake.
What belongs in it is the knowledge people reach for and cannot find:
- Decisions and the reasoning behind them. Not just that you chose a supplier, but why, what the alternatives were and what would make you reconsider.
- How things are done here. Onboarding, pricing rules, approval paths, the way you handle the awkward customer type.
- Who owns what. The person, not the department.
- What customers have said. Objections, requests, the phrase that made the deal close.
- What has been tried and failed. The most expensive knowledge a company has, and the first to be lost.
What does not belong: raw meeting transcripts, three versions of the same policy, anything without a date and an owner, and anything time-sensitive that will be wrong in a month.
Why it is not Notion, SharePoint or Copilot
This is where most knowledge projects go wrong, so it is worth being precise.
Notion, Confluence, SharePoint and Google Drive are places knowledge can live. Copilot, Gemini and ChatGPT are tools that can read those places. Neither of them decides what is true. If your pricing sheet exists in four versions, Copilot will cheerfully quote one of them, with confidence, and the person asking has no way of knowing it was the wrong one.
An AI assistant without a governed memory does not know your business. It guesses, fluently.
A second brain is the layer in between. It says: this is the current version, this person owns it, these people may see it, this is when it was last checked. It is what makes the tools you already pay for worth paying for.
Two ways to build one
In practice there are two architectures, and the right choice depends mostly on size and mess.
Curate at write time. Knowledge is written into a structured, interlinked set of documents, each with an owner, a date and a type. Think of a wiki that an AI helps you maintain, versioned in git so you can see what changed and when. Meta's engineering team described a version of this in September 2026: curated documents for the dense, frequently used knowledge, and raw retrieval only for the sparse reference material. Open formats such as the Open Knowledge Format are making this portable. For a company of twenty to a hundred people this is usually enough, it is cheap, and you own every byte.
Retrieve at read time. The system connects to your email, chat, drives, CRM and ticketing, indexes the lot and answers questions from whatever it finds, enforcing each person's permissions at the moment of retrieval. This is what enterprise search vendors sell. It is necessary when the sources are large and the permissions complex, and it is expensive to do well.
Most companies should start with the first and add the second only where a source is too large or too live to curate. Both can later be extended with a graph of people, projects and decisions when questions start needing more than one hop, and the vendor data suggests roughly eight in ten questions never do.
What keeps it alive
A second brain that nobody tends becomes a third brain: the old wiki everyone ignores. Three things prevent that.
- An owner per source. Not a committee. A name.
- A freshness rule. Anything older than its threshold gets flagged: update it, archive it or extend it. Thirty minutes a week.
- A question set. Fifty real questions the company should be able to answer. Test them when you change something. If the score drops, you broke something.
None of this is technically hard. All of it is organisationally hard, which is why it is the part we spend most time on.
Where to start
Do not start by buying a tool. Start by writing down the fifty questions, finding out where the answers live today, and checking who could see what if you connected an AI tomorrow. That is a two to three week piece of work, and it tells you whether you need a wiki, a retrieval system or just a tidier Drive.
Then build one slice, for one team, and measure it. The second slice is much easier.
Sources
- Organizational Second Brain: how AI learns from experts · Meta Engineering
- The AI Second Brain · Forte Labs
- Open Knowledge Format · okf.md
- Vector RAG vs GraphRAG · Neo4j