The AI never knew in the first place

When an AI tool does something different today than it did yesterday, the model is rarely the reason. The reason is what it was given this time. How reliable an AI setup is depends less on the model than on the text it runs on.

Every request starts from zero

A language model does not remember yesterday. The vendors say so plainly in their own documentation: each request stands on its own, with no state. A chat feels like a conversation because every new message sends the whole history along with it. The model reads all of it again, from the top, every time.

That text is called the context window, and it is everything the model knows about you at that moment. What is not in it, the model did not forget. It never knew.

Why it feels like forgetting

On a Tuesday you explain to a tool, at length, how your proposals should sound, which clients you are on first-name terms with, and which price is not up for negotiation. It does a good job. Two weeks later you open a new chat, and it writes like everyone else again. The obvious conclusion: the model is unreliable, time for a better one.

The better model gets the same blank start. What changed was not the intelligence but the note in front of it. Tuesday’s good answer came from an explanation that lived in that one chat. The chat is gone, and the explanation went with it.

The memory belongs to the tool

The big chat tools now have a memory, or projects with their own instructions. That helps, and it has two catches.

The memory files away, as you chat, whatever the tool decides matters. It is a selection the tool makes, not you. You can look at it and correct it. When did you last check?

And it lives in that one tool, often in one project inside it. The assistant in your email client does not know what you told the chat window, and neither does the automation that sorts your invoices. Three tools, three versions of your business, none of them complete, each wrong in its own slightly different way.

More text is not better

The obvious response is to pour everything in: every email, every set of notes, the whole manual. There is room for it. But a model does not read a long text evenly.

In 2023 a group of researchers led by Nelson Liu showed that language models use information at the beginning and the end of a long context best, and noticeably worse when it sits in the middle. The paper’s title says it: “Lost in the Middle”. One of the major vendors writes in its own documentation that accuracy and recall degrade as context grows, and that choosing what goes into the context matters as much as how much room there is.

So context is not a quantity. It is a selection. One page that is right beats an archive that has the right paragraph somewhere in it.

Write it once, let every tool read it

What goes in

Who you are and who you work for. The rules that are not negotiable. The decisions already made, each with one sentence on why, so the tool stops reopening them at every turn. And what a good result looks like, ideally as an example you liked. That should fit on one page. If it does not, you are not yet sure what you want yourself, which is worth knowing.

Where it lives

In a file you own, not in a vendor’s settings. From there it goes into the instructions of every tool you use. When something changes, you change the file and paste it in again. It is not elegant. It is also the simplest way to make every tool know the same business, rather than three rumours about it.

Along the way you end up with something you could use without any AI: a description of your business that does not live only in your head. Switch tools, and it comes with you.

The model is the replaceable part

Vendors replace their models regularly and retire the old ones. If you look for reliability in the model, you start over with every switch. If you look for it in the text the model runs on, it stays in your hands, and the next model gets it on day one.

So the question is rarely which model is best. It is what you actually told it.

Open a new chat in the tool you use most and ask: “What do you know about my business?” The answer is everything it works with. Write down what is missing or wrong. That is your first page.