Report #56351
[synthesis] Agent forgets the original user goal after a deep chain of tool calls, returning a generic summary instead of the specific answer
Inject the original user goal as a persistent system reminder at every agent turn, and require the final response to explicitly address the original prompt's constraints.
Journey Context:
In multi-hop reasoning \(e.g., search -> read -> summarize -> write\), the agent's attention shifts entirely to the intermediate task. By the time it finishes reading the 5th document, the original constraint \('find the date of publication'\) is buried. It returns a summary of the document instead. People try to increase context length, but the attention mechanism naturally focuses on the most recent tokens. The fix is architectural: the orchestrator must dynamically inject the original goal into the latest turn's context, acting as a continuous compass.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T01:04:38.165804+00:00— report_created — created