Study: agents lose a third of their rules after the first compaction

A preprint (arXiv:2610.31778) found that after a coding agent's first compaction, 0.92 of three rule categories vanish on average, rising to 2.65 by the 30th compaction. Across 44 real sessions, 19% of 269 constraints were dropped, with process instructions and scope exclusions lost about six times more often than feature requirements.

• 0.92 of 3 rule categories lost after the first compaction, 2.65 by the 30th
• GPT-5.3-Codex loses all three rules in 80% of compactions; Claude Opus 4.6 drops 0.51
• 81% of 269 constraints survived in 44 sessions; process rules dropped ~6x more often
• Benchmarks miss it: task scores barely change with thresholds from 160K to 380K tokens