Constraint Inflation: Context Compaction in LLM Agents Rarely Loses Instructions but Often Overstates Them

Abstract Long-running language-model agents compress their conversation history when it exceeds the context window. Prior work shows that this compaction can lose user instructions; we study the opposite failure, constraint inflation, in which a summary keeps an instruction but makes it broader or stronger than the user meant. We introduce InflateBench, 240 agent sessions with six planted items of known scope each, and compare seven compaction methods, including two published fixes for instruction loss, with two model families, model judges checked against human labels, and a behaviour test. All hypotheses were preregistered. Summaries almost never lost a standing instruction (below 0.8%), but inflated 13% of eligible items with a free summary and 19% with a sectioned template, mostly by turning preferences into requirements. One-offs without an explicit scope word were promoted to standing rules in 30% of cases, against 12% with one. The fixes for loss had no loss to prevent, yet raised inflation to 23% and 41% and kept replaced values as current in up to 94% of cases. Compaction with pinning made an agent repeat a one-off about six times as often as with the full history, though in only 4% of episodes.