Every business has a person everyone asks. How do we quote this. Where's the template. What did we tell that customer last time. When that person is you, the answers live in your head, and the team waits for you or guesses.
An AI assistant can answer those questions, but only from what it can read. The work is writing the answers down in a shape both a new hire and a model can use, keeping them in one place, and keeping them current.
Larger companies are doing this at scale. Notion built
Lore, open-source shared memory for agents stored as
Notion pages. On the hardest tasks in their test, the ones the agent never solved without memory, the agent with seeded memory recovered about 46% of the failures.
Notion's own caveat comes with it: the takeaway is not that memory magically fixes agents.
List the questions you answer more than once
For a week, every time someone asks you something you've answered before, write the question down. How do we price a rush job. What's the refund rule. Which supplier for which part. Who can approve a discount.
That list is the table of contents. A company brain that starts from a blank workspace fills up with whatever was easiest to paste. One that starts from the questions people ask fills up with what they need.
Notion's engineering team named the failure to avoid: "The first bad version of a memory system is a junk drawer with an API."
Checkpoint
A list of real questions, each one asked more than once.
Write one page per topic
the step that pays off
Take one topic from the list, talk through it into a voice memo or scribble notes, and run them through the prompt below. Every page comes out the same shape, with an owner and a date, and every gap is marked
[ASK: ...]for you to fill before it goes in.Keep the pages in one place the whole team and your AI tool can read: a
Notion workspace, a shared drive folder, or a project in
Claude with the pages added. Ten pages that are right beat a hundred that are nearly right.
Turn your notes on one topic into a brain page
Pro tip
Fill every [ASK] before the page goes in. An AI reading a page with a gap will answer the gap anyway.
Checkpoint
Ten pages in one shape, each with an owner and a date, no [ASK] left.
Decide who can read what
Before anything sensitive goes in, split the brain. Pay, personal details, customer payment details and anything under a contract go in a separate space with its own access, or stay out entirely.
Stripe's internal agent, Kai, is used by more than 10,000 employees every week, and Sharadh Krishnamurthy's account of building it treats "projects" there as a governance mechanism, not just a folder. The same idea works at any size: the space an AI can read is the space the people using it are allowed to read.
Checkpoint
A shared space everyone can read, and anything private kept out of it.
Ask it before you ask a person
Change the habit: questions go to the brain first. The prompt below makes the answer cite the page it came from and say "Not in the pages" when the pages are silent. That second rule is the useful one. Every "Not in the pages" is the next page to write.
Ask the brain, and get the page it came from
Checkpoint
Answers that name their page, and a list of the pages still missing.
Fix the page, not the answer
When the brain gives a wrong answer, don't correct it in the chat. Correct the page, change its "Last checked" date, and ask again. A correction in a chat is gone the next time anyone asks.
Once a month, the owner of each page reads it and either confirms the date or fixes it. Pages nobody owns are the ones that go stale first.
Checkpoint
A monthly check, with every page's date current.
The takeaway
You end up with a set of plain pages, each with an owner and a date, that answer the questions your team used to bring to you. Anyone can ask the brain and see which page the answer came from, and the questions it can't answer tell you what to write next.
The limit: the brain is only as current as its pages. It won't notice that a price changed or a supplier left. Someone has to change the page, and a page with no owner will drift. It also can't make a judgement call you haven't written down.
Variations
Onboarding from the brain. Give a new hire the page list on day one and have them ask the brain before asking a person. Their "Not in the pages" answers are a free audit.
Customer-facing answers. Once the pricing and policy pages are solid, the same pages can ground replies to customer enquiries. Keep internal-only pages out of that set.
One page per recurring decision. Who approves a discount, when you turn down a job, how you handle a late payer. Decisions written down once stop coming back to you.
Don't import everything you have. Dumping every old document into the brain is the junk drawer Notion warns about. Old versions and half-finished drafts will be read as current. Start from the questions.