Context compaction drops 83% of user rules
Long chats get summarized so the model can keep going. The summary keeps the job. It throws away how you asked to be treated.
Penn State researchers find that when AI systems compress a long conversation to free context, they retain only 17 percent of the session rules users actually set — things like "don't send email without my approval" or "confirm with me before you change anything." Zhiqi Wang, Yichi Zhang, Dongwon Lee, and Yuchen Yang call these session constraints: instructions that bind how the model behaves for the rest of the chat, not what the job is. Compactors are built for task continuity. They keep the goal, the current state, and the next step. A one-line "ask first" rule is not the task, so it dies in the summary. After that, the agent will send the email.
The team built CompInt, a 750-instance suite that injects those constraints into three long-context settings — multi-turn chat, agent trajectories, and long-horizon research — then checks whether the compacted memory still contains them. Most compactors they tested, from a keep-the-last-five-turns baseline to LLMLingua-2 and prompt-based summarizers on open models, do worse than leaving the full context alone. With the uncompressed conversation still in the window, models follow the rule 59 to 71 percent of the time, depending on the setting. After compaction, compliance often sits only a few points above the no-rule baseline. A prompt written specifically to preserve user constraints still keeps retention below 40 percent on the chat dataset. GPT-5.4-mini is the one model that sometimes beats the uncompressed baseline — and even that result is brittle, swinging from 98 percent retention on research traces to under 7 percent on WildChat with the same prompt family.
The useful finding is that this is not a prompt problem. The authors attach a small add-on, built on Qwen3.5-9B, that reads each user turn, lifts out session constraints into a separate list, and staples that list back onto the summary. No training, no change to the main model or the compactor. Retention jumps to 95.6 percent on agent traces, 95.1 percent on research tasks, and 90.3 percent on multi-turn chat. The paper and the evaluator are public.
Labs sold context windows as memory. Compaction is how those windows actually work once a session gets real, and the summarizer is optimized to keep the project moving, not to remember that you said no. "Don't delete until I confirm" is exactly the kind of instruction that is not the task and is the whole point. If your product auto-compacts a long thread, you already have this hole. The fix is a separate register for user rules, not a nicer summary prompt.
What to watch: whether the major agent runtimes start treating session constraints as first-class state, the way they already treat the system prompt.
If your agent compacted a long chat this week, did it still honor the one rule you actually cared about? Tell us in the comments.
Sources: Lost in Compaction (arXiv) · The Decoder · CompInt (GitHub)
