Title: The Paradox of Honesty: Deciphering Memory Confabulation in Reflexive Agents

Abstract

Reflexion-style agents utilize self-generated reflections as a form of memory, operating under the implicit premise that they can accurately identify their own shortcomings. However, our findings demonstrate that this premise is prone to systematic failure. Across the ALFWorld and HumanEval benchmarks, we observed that agents often encode and retain confident yet erroneous interpretations of the task. These agents persist in acting upon these flawed interpretations in subsequent trials, despite the environment resetting to the original, correct task parameters each time. We term this phenomenon "memory confabulation."

To address this, we introduce the Reflection Repetition Rate (RRR), a metric derived from logs that identifies instances where agents repeatedly depend on incorrect reflective content. Applying RRR, we pinpointed 16 "frozen" environments within ALFWorld—cases where none of the 121 reflections referenced the correct target object—and four similar instances in HumanEval. Our proposed mitigation strategy substitutes open-ended self-diagnosis with a programmatic extraction of trajectory-level failure signals. This approach significantly improved performance, raising the rate of correct object mentions from 0% to 86% and lowering the RRR from 0.64 to 0.10. Furthermore, it enabled the resolution of three of the 16 frozen ALFWorld environments. These results suggest that reflective memory may inadvertently reinforce false beliefs rather than facilitate their correction.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC