Why a Long Conversation Hurts the Answers
A conversation does not spoil with age. Length spoils it, and the reason is how memory works here.
The model has no memory
The model itself remembers nothing between messages. It processes every request from scratch, as if it were the first.
The memory is created by the application around it. With each new message, the app sends the model the entire previous conversation again: all your messages, all its answers, every attached file. The model reads all of it and answers the last message.
That is the price of a long conversation. The more you have said, the more text goes across the network with every reply.
Deleting a message from the window does not undo the fact that it was already sent to the server. You are removing it from future transmissions, not from past ones.
The limit on length
The context has a limit, and it is large: a document of several hundred pages fits in the window of a current model. Work conversations rarely reach that line.
A big window does not remove the limit. Answers get worse before the window fills up.
As a conversation grows, early details start to get crowded out, and the model leans on what was said later.
In practice it looks like this: at message forty, a condition you canceled at message ten comes back. Or an important detail from the beginning stops showing up in the answers.
Signs that it is time to stop: the model repeats itself, mixes up your earlier corrections, answers beside the question, or offers again what you already rejected.
How to move to a new chat
When a conversation has gotten too long, do not keep it going out of inertia.
You paste that summary as the first message of a new conversation. The conditions have moved; the old history stays in the old chat.
People do the same thing when they come back to a task a week later: starting from a short summary is easier than making the model read through 200 messages of an old discussion.