hesketh.ai
Write to me

I built one shared memory for a whole company.

It became the centre of its AI. By the time I left, the company had asked that it become the base of its AI work, and its engineers were splitting it into modules. The pattern I took with me now runs my own businesses.

The three parts, drawn as the documents they were Top, the part that listens every day: a single ruled sheet with a margin of marks, one for each source. Middle, the part that changes with the plan: a folder with a tab, the plan inside it. Bottom, the part that never changes: a bound volume in solid ink. It loaded first, every time.
03 THE PART THAT LISTENS every day
02 THE PART THAT CHANGES with the plan
01 THE PART THAT NEVER CHANGES loaded first, every time
Three layers, kept apart. The one that never changes always loads first.

WHAT GOES WRONG

Every AI tool starts from nothing. So people give it a snapshot of themselves. But a snapshot goes stale overnight. So people make more files, and soon they are looking after the system instead of using it.

Inside a company, it happens to everyone at once, and the work comes back sounding like whoever asked, not like the company, and not always keeping to its rules.

Decay is the enemy of accuracy, and of being able to trust the machine. A company needs a memory that stays current, and that anyone can ask.

WHAT I BUILT

  1. 01

    The part that never changes

    It taught the AI who the company was, before it did anything else. It held the mission, voice, products and strategy, the principles its leaders worked by, and the rules for using AI. It was not a prompt. It loaded first, every time. I used to say it gave the AI a university degree in the company. So every piece of work, from a marketing brief to an analyst’s code, sounded like the company and kept to its rules, whichever person or AI made it.

  2. 02

    The part that changes with the plan

    It told the AI what the company was trying to do in the months ahead, and it was swapped when the plan changed. Each plan was a large document, written for the AI to read. Tags marked what mattered most and asked the AI to say where each claim came from. It let the AI work as if it had lived through the current phase. Each department loaded the one for its own priorities. It beat decay without a database kept up to date by hand.

  3. 03

    The part that listens every day

    Every night it read what customers were saying in public, and by morning had turned hundreds of comments into a one-page report a leader could read before the day began. It read them through the part that never changes, so the company’s values and voice decided what mattered. The report came in the company’s look, every source linked: the day’s most important signal, what customers said about the product and its rivals, the mood, and each risk with the customers’ own words and a suggested action. An agent I built signed it as its own work. People read it, and people decided what to do.

  4. Before a meeting, people from two departments could both load the plan, and the AI would tell each one what the other side most needed. It stopped being a writing tool and helped people understand each other.

    The three ran as one loop. The daily part heard what was happening, the part that never changes judged it, and the plan said what to do about it. People got priorities that lined up. Then it ran again.

WHAT HAPPENED

I built it alone, with no team, no mandate and no budget. I had AI help all the way, but the ideas were mine.

Then other people built tools on the part that never changes. One helped people write marketing briefs, where most work starts, so the work began in the company’s voice and inside its rules. The value was time, and better work. The memory ended up under very different kinds of work: data analysts writing code, customer care sorting what came in, checks on software, and the leadership’s own office.

When I left, the base the company asked for was taking shape: one core for who the company is, then a package for each department.

The hard half was people. Leaders own their priorities, and they will not take them from a company AI. So a person stays in the loop, all the time. It helped that I was inside, where the data is to hand. From outside, access and confidentiality get in the way.

A snapshot from Monday is stale by Tuesday.

THE PATTERN I TOOK WITH ME

I call the pattern Substrate, Patch, Pulse, one name for each layer. The substrate answers: what never changes? The patch: what is true right now that wasn’t true last quarter? The pulse: what happened today? Each layer covers another’s failure. Merge them into one document and they start to decay. It came out of the builds, one at a time, not from theory. Then I did it over and over again. Here is where it has run.

Inside a company

The story on this page.

For my father

The first time it ran for someone else, it wrote my father’s letters, and it now sits under Hesketh & Son.

For myself

My own businesses run on it, through Forge, the set of AIs I built to work with.

What I build now starts from those three questions.

If you want this in your own company, it works best built inside your own walls, on your own data, with a person in the loop. Here is how I work with companies.

READ NEXT