An assistant that answered well for twenty exchanges starts forgetting things and contradicting itself. The cause is in how conversation is supplied to the model on each turn.
How a conversation is actually sent
Models do not retain a conversation between messages. Each reply is generated by sending the entire history back as input.
Every turn therefore costs more than the last, and the input grows until it approaches the context limit.
What feels like an ongoing relationship is a series of independent requests, each carrying the transcript with it. The continuity exists in the transcript, not in the model.
What happens when the limit approaches
Something has to be dropped, and the usual policy removes the oldest turns first.
Those early turns often contained the setup: the goal, the constraints and the corrections issued at the start.
Some products summarise the removed portion instead, which preserves the substance and loses the specifics, and the specifics were usually the operative part. None of this is announced, so the user sees only that the answers have changed.
Why instructions fade before facts
An instruction given once at the beginning competes against a growing volume of later text, and its relative weight falls with every exchange.
Facts stated repeatedly through the conversation survive, while a formatting rule mentioned once early does not.
This is why an assistant that respected a constraint for ten messages abandons it silently, without any event that would explain the change.
How errors get reinforced
Once a wrong statement enters the history, it is resent with every subsequent turn and functions as established context.
The model conditions on its own previous output, so a mistake becomes progressively harder to dislodge as the conversation builds on it.
Correcting it in place works less reliably than editing the earlier message or starting again, because the correction is one line against many reinforcing it.
What actually helps
Starting a new conversation for a new task is the single most effective habit, since it removes accumulated context that was never relevant.
Restating the important constraints periodically keeps them near the end of the input where they carry the most weight.
For anything long, keeping the durable requirements in a document that is pasted at the start of each fresh conversation is more reliable than expecting a chat thread to remember them.