Reflexion Agents Systematically Store Wrong Diagnoses — RRR Metric Detects It

Reflexion-style agents on ALFWorld and HumanEval repeatedly store confident but incorrect self-diagnoses as memory, persisting across environment resets — in 16 ALFWorld environments, 0 of 121 reflections named the correct target object. The paper introduces Reflection Repetition Rate (RRR), a log-based metric to detect this confabulation, and proposes replacing open-ended self-diagnosis with structured prompting as mitigation.