Attention budget

The attention budget is the finite amount of influence a token has to distribute over every other token in the context window. The weights across all of its attention relationships add up to a fixed total, so when the context grows, each token's budget is divided among more candidates. A short, focused request lets the important tokens spend most of their budget on each other; a bloated one spreads the same budget across thousands of bystanders.

This is why you can have a fact plainly in the window and still get an answer that ignores it. The instruction you gave in message three is still present, but by turn forty it is competing with a mountain of file contents and command output for the model's attention, and it loses a little on every request. The instruction was not forgotten; it was outvoted.

Spend the budget on purpose. Load fewer, more relevant files; summarise noisy output before it enters the transcript; restate the one rule that matters in the message where it matters, rather than trusting a rule from an hour ago to still carry weight. When a session has grown past the point where focus holds, clearing or compaction resets the budget.