AI Agents Break Rules They Still Remember After Context Compaction

ARCBench finds AI agents violate retained rules 2.6x more often after compaction; the strongest defense preserved rules in 99% of runs yet violations persisted at 29.4%.