Shared Recurrent Memory Improves Multi-agent Pathfinding

Multi-agent systems are gaining attention in AI research due to their ability to solve complex problems in a distributed manner, but coordinating multiple agents remains challenging. Inspired by the global workspace theory, we introduce the Shared Recurrent Memory Transformer (SRMT). SRMT extends memory transformers to multi-agent settings by pooling and globally broadcasting individual working memories, enabling agents to implicitly exchange information and coordinate behavior. We evaluate SRMT on the Partially Observable Multi-Agent Path Finding in a bottleneck navigation task requiring agents to pass through a narrow corridor. SRMT consistently outperforms a range of reinforcement learning baselines, especially under challenging reward structures with sparse feedback. Our experiments also demonstrate that SRMT generalizes effectively to environments with significantly longer corridors than those seen during training, highlighting its scalability and robustness. These results suggest that incorporating shared memory structures into transformer-based architectures can enhance coordination in decentralized multi-agent systems.