New AI Framework Prevents Multi-Agent Memory Bias

Multi-agent systems that continuously accumulate memories from various agents often face a critical issue called "Memory Correlation Bias." This bias occurs because existing memory arbitration methods typically treat retrieved memories as independent pieces of evidence, combining them through simple voting or weighting. However, memories from different agents might originate from the same upstream source or share a common bias, leading to correlated evidence being counted multiple times and creating a false majority in decision-making. To counter this, researchers propose the Correlation-Aware Memory Arbitration (CAMA) framework. CAMA is designed to jointly decouple retrieved memories and actively recover any missing independent evidence. It achieves this by modeling retrieved memories as query-conditioned evidence groups and combining neural dependency inference with provenance-based symbolic priors. This allows the system to estimate the effective number of independent evidence sources, thereby preventing correlated memories from skewing decisions. Furthermore, CAMA includes a sequential recovery policy that actively retrieves alternative evidence or traces upstream sources when critical independent evidence is absent from the initial retrieval set. This policy aims to gather sufficient independent evidence for reliable arbitration while minimizing retrieval costs. Experiments across multiple benchmarks demonstrate CAMA's superior performance, effectively suppressing false majorities caused by correlated memories.