How we think about company brains

A shared brain learns the company once, so every agent starts from the same view.

Before we built Cerebro, I would start a task for a client by telling Claude to pull context from our meeting transcripts, Gmail and Slack. It would read tens to hundreds of thousands of tokens of raw context to rebuild the history of our work with that client, so that it knew the current state of things before it could begin.

For the next task on the same client, it did the same thing again. And Claude was doing that same processing again for each of my teammates working on that client.

That creates two problems. The first is efficiency: you pay for the same processing on every task, for every person. The second is variance: agents read the same material and interpret it differently, so two people can act on different answers about the same project.

If it is just you working, you don’t really need a brain, because good connectors and a few skills are enough. The minute you have a team, or many teams, all directing agents, both problems show up. They grow with every person you add and every month of history.

The brain does that processing once, on behalf of everyone. We think of the raw sources as bronze: accurate, but unrefined. What the brain holds is gold: checked, organized, and traceable back to the meeting or thread it came from. Agents read the gold first and go back to the raw source only when they need to.

One client process in our brain draws on about 10 hours of meetings, five spreadsheets and ERP data, and it shows up as three steps.

A core job of the brain is to keep bad context out, so anything read from it can be trusted.

Context rots over time. Old decisions, half-right summaries and duplicates pile up, and every answer built on them gets worse.

It is easy to store information and easy for AI to search it, but it is hard to store only context that is accurate. Most companies have lived this with Notion, Confluence or Obsidian: everyone knows how much old material is in there, so nobody trusts what comes out of it.

Garbage in, garbage out is a cliché, and it is also what happens here. Claude or ChatGPT given messy or incomplete context produces slop, and someone then has to catch it.

The brain does its checking at the point of entry. A change is proposed and checked against what is already there, and most of those checks run on their own. The owner of that area creates and maintains the check, and reviews the exceptions and ambiguous cases.

Every area of the brain needs an owner. We think a brain that maintains itself is going to be wrong for a long time. If you are the CFO, someone on your team should own the definitions of your financial metrics, because everyone in the company uses them and your team has the authority over them. In the same way, engineering owns how the data should be queried.

Every change is logged, so you can see what the brain held on any date and why. When a decision is reversed, both versions stay in the log, each with its date.

Anyone building on top of the brain can skip the checking, because it was done on the way in.

A brain exists to help you adapt to your environment, and every company operates in a different one.

Most brains share a few parts: a decision log, the definitions of your metrics, and a wiki of your people, customers and products.

The real advantage comes when you tailor the brain to the shape of the work your company does. We are a services business, so Cerebro separates client knowledge from internal Glia knowledge. Each client’s knowledge is organized along three dimensions: people, process and systems. Each client also has its own wiki, which tracks concepts and definitions such as a directory of their people. Most of our work starts from the same context, which is how that client operates today, and our agents need it many times a day.

It also let us do something we could barely do before, which is to show a client a map of their own company, from the top level down to a single step. We had all the raw context for it long before, but it was scattered across transcripts and files.

To find your shape, ask what your team’s agents need to understand over and over, across many people. The answer is different for every company, and it changes as the company does. That is why we don’t believe a brain can be bought purely off the shelf.

We treat this as build and buy. Buy the software that acts as a sensor, meaning it collects data that would not exist anywhere otherwise: your notetaker, your database, your chat tool. Build the parts that need judgment specific to you, such as what counts as a decision, how you define your metrics, and what is worth keeping.

Keep that context somewhere you own and can move, the same way you would not let one vendor own your database.

A brain is built one pathway at a time, and every pathway has the same three steps.

Every pathway follows the same pattern:

Start with meetings and decisions. Meetings are the richest source most companies have, and every team we have worked with needs to know what was decided. Notetaker summaries are full of false positives, so the process step matters from the first day.

Then repeat the pattern for the next source. Decisions get made in meetings and overturned in Slack, and you only see that once both pathways exist.

Past a certain size, the brain needs infrastructure and an owner. Cerebro runs on servers and has a dedicated engineer.

On November 19, we are running a 90-minute public workshop, How to build a brain for your team. You leave with a working decision log built from your own meetings, and the pattern for the next pathway.

There are 20 seats at $99 each, and all proceeds go to a nonprofit that builds AI fluency in underserved communities.