Comment by Bolwin

You're entirely correct in that it's a different model with every message, every token. There's no past memory for it to reference.

That said it can still be useful because you have a some weird behavior and 199k tokens of context, with no idea where the info is that's nudging it to do the weird thing.

In this case you can think of it less as "why did you do this?" And more "what references to doing this exist in this pile of files and instructions?"