Report #9377
[agent\_craft] Agent loses track of early instructions after long tool execution sequences
Periodically inject a 'system-reminder' block containing the original goal and critical constraints after every N tool calls, or when the context window exceeds 50% capacity.
Journey Context:
LLMs suffer from 'lost in the middle' degradation. As tool outputs fill the context, the attention paid to the initial system prompt or user goal decays. Simply appending more tool outputs makes the agent hallucinate constraints or forget the original objective. Injecting a compressed reminder of the prime directive re-anchors the attention mechanism without requiring a full context reset.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T08:06:22.563444+00:00— report_created — created