When Memory Is Not Enough: What Drives Robustness in AI Decision Systems? Evidence from Sequential Financial Decision-Making

AI systems are increasingly used to support repeated decisions in environments where conditions change over time. Many of these systems combine large language models, retrieved experience, and rule-based execution layers, but their behavior is often evaluated as a single end-to-end system. As a result, when such systems become more stable or risk-aware, it remains unclear whether the improvement comes from model reasoning, relevant memory, stochasticity, or the surrounding control logic. This paper investigates this question in a sequential financial decision setting using Apple, Tesla, and NVIDIA over 2021–2023, where memory access, coordinator choice, retrieval quality, stochasticity, and execution-module configurations are examined through targeted comparisons. The analysis suggests that memory can provide useful context but does not by itself produce robust behavior under adverse or shifting conditions. The full configuration exhibits stronger risk-adjusted and downside behavior than configurations that remove memory or use retrieved experience without model-based synthesis. The strongest observed robustness occurs when relevant memory is combined with bounded model-based coordination and constrained execution. Within this financial testbed, the findings suggest that retrieval relevance, bounded model-based coordination, and execution constraints warrant joint evaluation rather than attributing observed robustness to coordinator choice alone.