ReM-MoA Sustains Multi-Agent LLM Scaling with Reasoning Memory

Mixture-of-Agents (MoA) architectures aim to enhance large language model (LLM) inference by arranging multiple agents in layered reasoning pipelines. However, a significant challenge has been the inability of existing MoA variants to maintain performance improvements as the depth of these pipelines increases, often leading to degradation or early saturation. Researchers have introduced ReM-MoA, a novel memory-augmented MoA framework designed to overcome these scaling limitations. ReM-MoA incorporates two key mechanisms: first, a "Ranked Reasoning Memory" that persistently stores and ranks reasoning traces from all layers using a comparative Reviewer Agent. Second, a "Curated Diversified Memory Routing" scheme exposes different agents to distinct combinations of successful and failed traces. This dual approach ensures the propagation of high-quality reasoning while preserving exploration diversity. An optional multi-domain Reviewer distillation pipeline further refines ranking quality through supervision from frontier models. Across five diverse reasoning benchmarks—including math, formal logic, code, knowledge, and commonsense—ReM-MoA consistently outperformed previous MoA variants. Its advantage notably widened with increased pipeline depth and width, establishing structured cross-layer reasoning memory as a critical component for scalable multi-agent inference.