Why Your AI Forgets, and What Fixes It

You explain your situation. You get a good answer. You come back two days later and explain the whole thing again.

Almost everyone using a hosted assistant seriously has hit this, and almost everyone assumes they are using it wrong. They are not. It is how the thing is built.

Three different kinds of forgetting

They get confused with each other, and they have different fixes.

Forgetting within a conversation. Long exchanges drift. Something agreed near the start quietly stops being applied. This happens because a model can only hold so much at once, and when the conversation outgrows that, the earliest part falls out of view.

Forgetting between conversations. A new session starts blank. Everything you established last week is gone. This is deliberate: each conversation is designed to be separate, partly for privacy and partly because storing everything for everyone would be enormous.

Never having known. The most common and least noticed. It cannot tell you what the roof work cost because it has never seen your accounts. No amount of conversation fixes this, because the material was never there.

Why the usual workarounds disappoint

People reach for the same three things.

Longer context. Newer models hold more at once, which genuinely helps the first kind of forgetting. It does nothing for the second or third, and pasting your life into every conversation is not a way to live.

Memory features. Hosted assistants increasingly remember facts about you between sessions. Useful, and worth turning on. But these hold a summary — preferences and a few facts — not your archive. And the summary lives on their infrastructure, which is exactly the thing some people were avoiding.

Uploading documents each time. Works for one document and collapses at scale. Nobody re-uploads a decade of paperwork to ask what something cost.

What actually fixes each one

The third kind — never having known — is the one that matters most, and the only fix is for the assistant to have read your material.

That means the material has to be somewhere it can be read, which is why this turns into a question about where the computer is rather than which assistant you chose. When your documents sit on a machine in your building, and the assistant runs on that machine, the third kind of forgetting stops existing. It has read the archive. It does not need to be told again.

The second kind — starting blank — stops mattering for the same reason. The context that made your last conversation useful was mostly facts about your world, and those are now permanently available rather than re-supplied by you each time.

The first kind — drift within a long exchange — does not disappear. It is a property of how models work, not of where they run. What changes is the cost of it: when something drifts, you can say “check the contract” instead of re-explaining the contract.

The honest limit

An assistant that has read your archive can still be wrong about it. Reading is not the same as understanding, and an archive that contradicts itself produces an assistant that contradicts itself. Memory is not the same thing as accuracy, and anything that claims otherwise is selling.

What you should expect is narrower and more useful: it should not need to be told what it has already been shown, and when it gives you a specific figure it should be able to show you where that figure came from.

A test worth running

Whatever you are using now, ask it something that requires knowing your own affairs. Not a general question — a specific one, with a checkable answer, about your household or your practice.

If it asks you to supply the information first, it has never known. That is the gap, and no prompt closes it.

Personalisation without sending anything away →

Find out what your own archive supports →