Context Inertia: agents won't let go of dead-end hypotheses — the fix is throwing out context

dbreunig · x · 2026-10-15

Drew Breunig names a new context failure mode, Context Inertia: during long-horizon tasks, a model that has committed to a hypothesis, subtask, or goal in its context keeps pursuing it even after evidence says stop or change course. It sits between context poisoning (an error enters context and gets repeatedly referenced) and context distraction (over-focusing on an overlong context), amplified by frontier models trained to be relentless instruction followers.

Anthropic's case study: the Memento harness from its self-replication research addresses exactly this — one Qwen agent debugged package versioning for hours when the real bug was an endpoint incompatible with the grader API. In Memento, each agent runs until a fixed timeout (1.5 hours in one experiment), writes a short note, and a fresh agent inherits only the codebase, the environment, and the note — nothing else.

A paper the same week, "Fork-and-Flush" by Jack Ziyang Cai and Christos Ziakas, hits the same note: periodically fork and flush the context rather than letting one agent carry it all. Takeaway: for long-horizon agents, forgetting is a feature, not a bug.

Original post →

Context Inertia: agents won't let go of dead-end hypotheses — the fix is throwing out context

More from coding & agent