When Context Won’t Let Go Residual Contextual Baggage and the Dethreading Problem in AI-Assisted Systems Building

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This is a conceptual paper, not an experiment. It proposes a name and a framework for something I kept running into while building software with AI models over long sessions: a line of reasoning that was useful at one stage seems to keep shaping the work after it stops being relevant, including after the model has explicitly said it moved on and stopped mentioning it. I call the continuing influence residual contextual baggage, and I call the difficulty of shedding it while keeping the underlying information dethreading failure. The distinction the paper turns on is between retention and influence. A model can hold information without that information deserving equal weight on every problem that follows, and my hypothesis is that current long-context systems may not separate those two cleanly.