Long AI conversations hit a wall when the context window fills up. The standard fix, called compaction, summarizes chat history to reclaim space. But researchers at Penn State have now quantified what gets lost in the process, and the answer is worse than most users assume: the rules you set for how the AI should behave.
These ‘session constraints’ are things like “confirm before making changes” or “never use my name.” They aren’t task instructions, so compression systems built to preserve goals and next steps simply discard them. The team built an evaluation suite called COMPINT to measure the damage. On average, only 17 percent of injected constraints survive compaction. Even with uncompressed context, rule compliance sits between 59 and 71 percent. After compression, compliance collapses to near the level of no constraint at all.
A targeted compression prompt helps, but not much. The researchers say retention stays below 40 percent even when the prompt explicitly asks to preserve user rules. That’s a security problem, not just a quality one. An agent told not to send emails without approval might do exactly that after compaction. Unauthorized tool calls and skipped verification steps become real risks.
A plug-and-play fix exists, though. The team built a small extractor module on Qwen3.5-9B that scans user inputs for session constraints and stores them in a separate list. When compaction happens, the module re-appends that list to the summary. Retention jumps to over 90 percent across three tested scenarios, including 95.6 percent for agent trajectories. No training required, no changes to the compression system itself. Both COMPINT and the extractor are on GitHub. This is the same tension we’ve seen with longer context windows in general: more capacity doesn’t solve the problem if the systems managing that capacity quietly drop what the user actually asked for.
