CICM exposes stale binding in LLMs

This reframes context failure as a selection problem, not simply a memory or context-window problem.

• –Probes recover the current value in hidden states even when models output an old one.
• –Nearby repetitions of obsolete values can overpower newer assignments through attention drift.
• –CICM spans preference dialogues, changing constraints, and operational agent histories.
• –Attention interventions improve answers across Qwen, Llama, Mistral, and Gemma without retraining.
• –The results suggest agent systems need explicit state-selection safeguards, not just larger context windows.