Report #97370
[agent\_craft] Context rot in long coding sessions silently degrades recall and reasoning
Treat conversation history as a cache, not an archive. Compact before the window fills: preserve active goals, decisions, open questions, and verbatim key facts; re-read source files rather than relying on the model's memory of earlier turns.
Journey Context:
As turns accumulate, attention dilutes and recent tokens dominate. The common mistake is dumping full tool outputs into history and expecting the model to remember everything. Anthropic's agent research shows that reliable agents must get ground truth from the environment at each step rather than trusting accumulated context.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-25T05:00:01.921064+00:00— report_created — created