Your agent is analyzing a complex 200-page commercial contract. By turn 30, it's giving inconsistent answers — saying "the indemnity clause is in section 4.2" in one turn, and "section 5.1" three turns later. The actual section locations haven't moved. Other facts (effective date, termination notice period) are also drifting between turns.
What's the architectural fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Scratchpad files are the structural fix for context degradation in extended exploration. Once findings are written to disk, they're no longer dependent on conversation-history attention — the agent reads them fresh each time it needs them. This is a recurring pattern in long-running analytical work: don't rely on the model to recall facts from buried turns; have it record findings as they're made and consult them as needed.
Lowering temperature reduces output variance but doesn't fix attention degradation in long contexts. The clauses haven't moved; the model is forgetting where it found them, which is a different mechanism from making things up.
Larger context windows let more information fit, but they don't solve the lost-in-the-middle effect — facts buried in the middle of a long input are still under-attended. Capacity isn't the bottleneck; structured persistence is.
Re-injection requires you to know which facts will turn out to matter, and burns tokens on every call. A scratchpad is the agent's structured memory; it scales with what's been discovered, not with what the user remembers to remind it of.