New Research Finds Optimal Memory Depth for Robust Consensus Networks

A new paper from Jiamin Wang, Jian Liu, Feng Xiao, Haibin Duan, and Yuanshi Zheng explores how local memory can enhance the robustness of consensus networks—systems where multiple agents (e.g., drones, sensors, robots) must agree on a common value without central control. The team proposes a protocol that uses linear extrapolation with single-step memory and a tunable memory depth, inspired by hierarchical temporal memory structures in neuroscience. They derive the necessary and sufficient condition for achieving consensus and prove an inheritable property: if a protocol works at one depth, it works at all deeper depths.

The key contribution is an analytical expression for the H₂ performance metric (which quantifies network robustness to disturbances) as a function of memory factor, memory depth, coupling gain, and Laplacian spectrum. Counterintuitively, the optimal memory depth is not intermediate but at either extreme—most recent or most remote—depending on parameter regions. This means that for a given network topology and gain, engineers can select the memory depth that maximizes disturbance rejection. The findings have broad implications for designing resilient multi-agent systems in autonomous vehicles, power grids, and IoT networks, where small perturbations can cascade into system failures.