Context Injection and Long-Horizon Reliability: Separating Duplicate Recognition, Local Processing, and Counter Tracking

The study scores long-horizon failures by breaking down a ledger task into three steps: recognizing seen records, local processing, and tracking a counter. It found that irrelevant context strongly disrupts counter tracking more than local processing. An explicit external reminder can restore the counter value even when context is disruptive.

Open the full episode

More from AI Radio