Team0
ProductHow it worksWorks withUse casesAppsPilotsBook a call
Building an agentic brain

Part 11

Teaching it your industry

Teach your brain one industry with a short list of the details that matter there, and test it on material it has never seen.

Yogev Ben-Tov · Published Nov 18, 2026 · 8 min read

A general brain knows who people are, who they work for and what was promised. That is not enough for a travel agency or an accounting firm. A travel agent asks "what does this traveller always want?" An accountant asks "when does this client file, and in which system are their books?"

To a general brain, "window seat as always" is one more sentence in one more email. To a travel agent it is the thing to remember about that customer.

We taught a brain two industries, travel and accounting. This part shows how: a short list of the details that matter in the industry, and four places where the brain uses that list. Then how to test it, and what the test will show you about the brain itself.

Start with the questions an operator asks

Sit with someone who does the job, or read a week of their mail. Write ten questions they ask about a client, the way they would ask them:

  • "What should I know before booking for this traveller?"
  • "Who approves trips at this company, and what is their policy?"
  • "What did we quote them, and is it still open?"

For each one, write the details a complete answer must contain, and anything it must not contain: another client's details, a value that was replaced.

Write the list

The list names the kinds of detail that stay true about a client until someone changes them.

Ours was 7 lines for travel: preferred airline, seat, meal requirement, direct flights only, preferred hotel, loyalty programs, the company's travel policy. For accounting it was 14, with lines like filing cycle, bookkeeping system, fiscal year end and what needs the client's sign-off.

You name the kind of detail, never its values. One line says "preferred hotel, one per city". Which hotels, which cities and whose they are all come from the mail.

Each line has four things and one sentence:

  • A name, such as seat_preference.
  • Who holds it: a person or a company.
  • One value or many. A traveller has one seat preference and several loyalty programs.
  • A scope, if it has one. A preferred hotel is per city.
  • What it means, in the industry's words. This sentence is for the model.

Leave out anything that is true once: this trip's seat, this quarter's late filing.

Keep the list as data your system loads, one file or table per industry, so a business outside that industry never sees it. An MD-based memory can keep it as a short file. A database can keep it as a table.

Use the list in four places

The list does nothing by itself. It teaches the brain because four steps you already built start using it.

1. Reading. When your system reads a source, show the model the list and let it decide whether a sentence states one of these details. This is how it catches a detail said in passing: "Book her to Lisbon, window seat as always" holds a booking and a lasting preference. Do not match on words. "Aisle" in a sentence about a supermarket is not a seat preference.

2. Storing. Store the detail under its name, on the record of whoever holds it: the traveller, not the trip; the company, not the person who mentioned its policy. Keep the scope with it, so the hotel in Lisbon and the hotel in Vienna are two values, not one that keeps changing.

3. Changing. A detail with one value has one current value. When the traveller who always wanted an aisle seat asks for a window, the window replaces the aisle, and the aisle stays in the history with its source. A detail with many values adds. The booking ends when the trip is over; the preference does not end with it.

4. Answering. When someone asks about a client, put that client's standing details first, each with its source. If you built a page in part 10, a client profile is the natural first page for an industry.

Each step existed before. The list tells each one what to look for.

The list is a seed, not a fence

The list does not limit what the brain learns. A detail that is not on it is still read and stored like any other sentence. What it misses is the special handling: one current value, a history, first place in the answer.

So do not try to write a complete list. Write the handful of details an operator would name without thinking, so the brain is useful on the first day, before it has seen enough mail to notice patterns.

Then let the brain grow the list. When it reads a lasting detail that fits no line, such as "always needs a child seat in the car", let it propose a new kind. Let it use that kind once it has seen it for several clients, and show a person the proposals so they can confirm, rename or remove them. A list the brain keeps growing from the client's own material is the goal. A list written by hand is how you start.

Expect the industry to test your brain

A new industry brings situations your earlier tests did not have. On the travel mailbox we found weak spots that had nothing to do with travel. We fixed each one in the brain, and accounting got the fix without any work.

Check these on your own prototype:

  • Who a detail is about. An assistant writing for her manager is the speaker. The preferences are the manager's. A first name alone is a hint, not a person.
  • A vaguer sentence replacing a precise one. "Economy on short flights" should not overwrite "economy under five hours".
  • Things that open and close. A quote, a provisional hold, a revised booking; in accounting, a filing or a missing document. A group that shrinks from 16 people to 14 has one current arrangement. A declined quote is closed, and still worth reading later as "what we offered".
  • A date without a year. "Due by 10 May", written in October, means next May.
  • Answers that wander. An answer about one client should not carry lines about other clients.

If a fix names the industry, it is a patch. Fix it so that the second industry gets it for free.

Test on a mailbox you have never seen

Your own test mail is practice. You wrote it, and you fix what you see.

Have a different agent, or a colleague, write an invented mailbox for the industry: 40 to 50 messages, several clients, changed preferences, an assistant writing for a manager, two people with the same first name. They also write the ten questions and the required details, and keep that key from you. They grade the answers. You see counts.

Score per detail. An answer that has the core right and misses two details should read as "5 of 7", not as right or wrong.

Then run a second industry before you write its list. Where answers got cleaner, your fixes were general. Where details are still missing, that is what its list is for.

Find where the details are lost

A score does not tell you what to fix. For every missing detail, check what your system stored.

When we did this, almost every missing detail had been read and stored. About a third were stored and not shown. About a third were closed too early. About a third were stored incompletely. Almost none were never stored, and none were on the wrong client.

So reading was not the problem. Closing and showing were. Both can be checked on the stored data with no model calls, so fix them there before paying for another run.

One caution: two runs of the same system do not give the same score. Run twice before you trust a small gain.

Measure

Add the industry questions to your set and keep the unseen mailbox as held-out.

python3 score.py questions.jsonl part9.jsonl part11.jsonl --out part11-scores.json

Compare per-detail counts, wrong answers, and lines about other clients. Then decide whether the next step is a line in the list or a fix in the brain.


Build and measure your own setup with the companion skill. See the series. For a maintained company install, talk to Team0.

Part 10Showing what it knows

About the author

Yogev Ben-Tov

Founder of Team0

REV 2026-11-18

Team0

A brain for each person and one for the company, read by every agent.

Product
  • Product
  • How it works
  • Sources
  • Walk through an example
  • Use cases
  • Apps
  • How a pilot works
Technology
  • The World Model
  • Architecture
  • Data model
  • Entity resolution
  • Truth maintenance
  • Trust and scope
  • Understanding Engine
  • Comparisons
  • The 31 problems
Second brain
  • What is a second brain
  • Markdown vs Team0
  • Memory vs understanding
  • How agents connect
  • A second brain you have to feed
Agents
  • Works with
  • Developers
  • Connect over MCP
Trust
  • Trust and control
  • Security
  • Privacy
  • Terms
Company
  • About
  • Blog
  • FAQ
  • Support
© 2026 Team0Every agent in your company, in the know.