Report #7705
[agent\_craft] Agent forgets initial system prompt or tool schemas after long context accumulation
Pin critical instructions \(tool schemas, core rules\) at the absolute beginning and end of the context window, or inject them dynamically right before the generation step.
Journey Context:
LLMs suffer from 'lost in the middle' degradation where attention weights dip for information in the center of a long prompt. Naively appending tool results pushes critical instructions into this attention dead zone. Moving core constraints to the prompt boundaries or injecting them just-in-time ensures they receive high attention weights.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T03:35:25.486555+00:00— report_created — created