One conversation is one workspace
A conversation with an AI is one workspace. Everything in it counts: your first question, the text you pasted, every answer. That space has a hard limit. When it fills up, the beginning drops out. The answers get worse, but you get no warning. Anthropic, one of the makers, calls this "context rot" itself. Research from 2025 showed that every model tested performed on average 39 percent worse in a long conversation than on a single question. That is not a gradual decline you notice easily. It happens quietly. There is a second mechanism too. When you hit your daily limit, the tool sometimes switches to a smaller, faster model. That model is not in the menu, so you do not see the switch. You think you are working with the same model, but you no longer are.
Do this: Start a new conversation per job, not per day. Moving from writing up minutes to comparing suppliers? New conversation. Otherwise details from one piece keep reading along in the next. Notice the tool forgetting agreements you made earlier in the same conversation? Then the workspace is full.
This is a real lesson, not an example written for this page. It was written during the send of edition 202636 on 1 September 2026, for the ‘diepgaand’ level, and translated here. The examples in a lesson follow the reader’s own field, so your team gets them in their own words.