Compaction study: agent constraints vanish 6x more often than task goals in long sessions
maksym_andr · x · 2026-10-31
maksymandr notes his team's context compaction paper is getting attention in Japan. The finding: as AI sessions grow long, models preferentially prune information deemed non-essential, and verbal instructions tend to disappear first.
- Experiments had Claude, GPT, and Gemini run long tasks while measuring which message content survived compaction
- "Goals" and "progress" persist regardless of session length
- "Rules/constraints" instructions are forgotten over time, even when told to preserve them
- Directive-style instructions ("do it this way", "don't do that") vanish 6x more often than "what to build" instructions
Practical takeaway: put rules you can't afford to lose in CLAUDE.md or AGENTS.md, which sit outside the compression scope.
Related event: Study Finds Context Compaction Spares Benchmarks but Drops User Constraints(4 posts)→
More from coding & agent
- Human-in-the-loop with Claude: rubber-stamping 400-line bash scripts as 'approval' — generativist · 2026-10-31
- How to actually evaluate a GraphRAG stack: multi-hop questions, conflicting facts, citations — JeremyCMorgan · 2026-10-31
- Suno ships MCP Beta, dev uses it to sync motion graphics to AI-generated music — enggirlfriend · 2026-10-31
- Count document passes, not tokens: batching questions cuts LLM scoring cost — petrusenko_max · 2026-10-31
- Open-source framework uses a second agent to catch AI slop before it hits your repo — kwerkee · 2026-10-31
- MCP-Powered Blender Demo Stuns With Realistic Metal Reflections — sidahuj · 2026-10-31