Every time a lab ships a longer context window somebody announces that memory is solved. It is a reasonable thing to think and it hasn’t been true yet.
Context is what the model is holding right now. It is enormous, it is impressive, and it is gone the moment the session closes. A bigger window means a model can hold more at once. Holding isn’t remembering.
The expensive thing is not the code
Code survives. It is on disk, anyone can read it again.
What disappears is everything around it. Why this approach and not the obvious one. What was tried in June and abandoned. Which part of the system looks wrong to every new engineer and has to stay that way for a reason that is not written anywhere. The hour your team spent arguing about a schema and the conclusion they reached.
None of that is in the repository. It lived in a conversation, and the conversation ended.
So the next session opens cold, proposes the approach that was already rejected, and somebody on your team explains it again. They have explained it four times. They will explain it next week.
Your engineers are already doing this work. They are the memory layer. They carry the context between sessions in their heads and re-type it by hand, and because it feels like just talking to the tool, nobody counts it as cost.
Retrieval gets you part of the way
The usual answer is to put everything in a database and search it when it seems relevant. That genuinely helps and we do it.
It is not sufficient, and the reason is a bit subtle. Most of what matters here was never written as a document. It was a decision, and a decision has parts: who made it, what got rejected, what would have to be true for it to be revisited. Flatten that into a paragraph of text to match against and you keep the words and lose the structure that made it worth keeping.
The bar we hold ourselves to
A cold session should be able to tell you what the team already settled, before it proposes the thing that was settled against. Not surface a vaguely related paragraph. Tell you the decision, and who made it.
We are not all the way there. It is the thing we work on most.