A conversation does not fill the window smoothly; it fills in blocks, one per turn, and every block has the same shape. Notice how few of each turn's tokens are the user's and how much is the model's reply: the agent spends most of its window on what it says, not on what it is asked. The system prompt is written once and never repeats, so after enough turns it stops mattering, while the turns keep coming at the same size each time.
Paid members: the worksheet and its printable PDF are below ↓