Long‑term multi‑agent systems continuously accumulate memories generated by different agents. Existing memory methods usually treat retrieved memories as independent evidence and combine them by voting or weighting. This independence assumption often breaks down in multi‑agent settings: memories written by different agents may share the same upstream source or bias, causing correlated evidence to be counted repeatedly and creating a false majority. We call this failure mode \textit{Memory Correlation Bias}.
To tackle this issue, we propose the Correlation‑Aware Memory Arbitration (CAMA) framework. CAMA models retrieved memories as query‑conditioned evidence groups and combines neural dependency inference with provenance‑based symbolic priors to estimate the effective number of independent evidence sources, thereby preventing correlated memories from forming a false majority.
Because critical independent evidence may be absent from the initial retrieval set, CAMA also learns a sequential recovery policy that actively retrieves alternative evidence or traces upstream sources before making the final decision, aiming to recover sufficient independent evidence for reliable arbitration while minimizing retrieval cost.
Experiments on several benchmarks demonstrate that, compared with state‑of‑the‑art baselines, CAMA significantly suppresses false majorities induced by correlated memories and improves arbitration accuracy.
Blogger's Review: By explicitly modeling the provenance of memories, CAMA turns “correlation” into a quantifiable prior, addressing a often‑overlooked evidence dependency problem in multi‑agent systems. Its sequential recovery strategy is both practical and elegant, making it a promising component for real‑world collaborative AI.