Why AI assistants forget decisions

Every AI assistant demo looks impressive for about five minutes. Then reality starts.

Something gets decided. Something changes. Something that was true an hour ago is no longer true. And the assistant — confidently — acts on the old version, because it never really knew the new one.

It's not a model problem

The reflex is to blame the model: it's not smart enough, the context window is too small, we need a bigger model. But watch where assistants actually break, and it's rarely reasoning. It's memory of state: what's already done, what's been decided, what's still open, what's been superseded.

A model that re-reads the entire conversation every turn isn't remembering. It's re-deriving — guessing the current state from a transcript, every single time. That works until two messages contradict each other. Then it has no principled way to know which one still holds.

Longer prompts don't fix this

Stuffing more history into the context is treating a state problem as a text problem. The transcript grows, costs rise, and the assistant still has no explicit notion of "this decision replaced that one." You've made the haystack bigger, not the needle easier to find.

What I think the shape of the fix is

Make state and evidence first-class, separate from the chat log:

• A decision is a record, not a sentence buried in history.
• When something is superseded, you update the record — you don't append a contradiction and hope.
• The model reads the current state, plus the evidence behind it, instead of reconstructing it from scratch.

The chat becomes the interface, not the memory.

What I'm watching for

If this is right, the assistants that feel reliable over long sessions won't be the ones with the biggest models — they'll be the ones that treat state as something you maintain, not something you re-guess. That's the thing I keep testing.

This isn't really about AI

Strip away the assistant and the lesson is the oldest one in state management, wearing a new costume: keep a single source of truth, model decisions as records you update rather than a log you keep appending to, and read the current state instead of re-deriving it from history every time. We learned this building UIs that sync across tabs and services that can't trust their own event log. Assistants just made the cost of ignoring it obvious again — and on a surface where everyone can see it fail.