Users report that a long-running research multi-agent loses early constraints (budget cap and excluded vendors) after many tool turns, then invents conflicting recommendations. Token usage per session is near the model limit. Which diagnosis best explains the failure mode?
Select an answer to reveal the explanation.
Short Explanation
You're seeing sliding-window amnesia: when sessions get long and token usage hugs the context limit, the model drops earlier turns. Budget caps and excluded vendors that lived only in early messages vanish, so later recommendations conflict. Fix it with durable constraint memory, summary or compaction that preserves entities and rules, and checks that re-inject hard constraints before finalization.
Full Explanation
Sliding-window amnesia is a documented context-window failure mode: as tool results accumulate, earlier tokens fall out of the active window, causing loss of constraints and entity continuity. The scenario explicitly cites long multi-turn tool use near the model limit and invented conflicting recommendations, which is consistent with amnesia rather than infrastructure side issues. The fix is not more GPU capacity or Front Door DNS changes; it is durable constraint memory, summary or compaction strategies that preserve entities and rules, and checks that re-inject hard constraints before finalization. The embedding-cluster GPU quota option would typically surface as retrieval failures, not silent constraint loss mid-session. The Azure Front Door custom-domain DNS misconfiguration option affects reachability, not conversational memory. The incorrect RBAC option on the storage account used for static marketing assets is orthogonal to agent context. Diagnose context loss with traces that show which turns remain in the prompt when the bad answer is produced. Remediation includes long-term or session memory for hard constraints, careful compaction that preserves critical entities, and evaluation tests that assert constraint retention after many tool turns.