Context Compactors Silently Drop 83% of Standing Rules, New COMPINT Suite Reveals

dair_ai · x · 2026-08-13

Current AI agents commonly use context compactors to save tokens during long conversations, but new research exposes a critical flaw: existing compactors silently drop up to 83% of "Session Constraints."

Session Constraints are standing instructions given by users (e.g., "do not delete any emails until I confirm"). The researchers introduced COMPINT, an evaluation suite testing multi-turn chat, agentic trajectories, and long-horizon research. Results show that compaction often leaves tasks worse off than running without compaction at all.

Retention swings based on structural factors like the compactor, prompt, context length, and injection location. Fortunately, the fix is straightforward: running an SC-aware extractor alongside the compactor recovers over 90% of retention without altering the base model or compactor.

Original post →

More from coding & agent

coding & agent channel →