David González-Martínez, Philipp Nazari, T. Konstantin Rusch, Shiwei Liu
arXiv 2610.18809 34 pages Read the paper
Layer outputs need a route to influence later layers; Depth Memory describes that flow as streams that layers read with queries and update with keys and outputs.
Layer mixing must retain useful associations without storing every earlier representation; Linear Depth Memory uses recursive least squares to keep a fixed-size regression fit that later layers can query.
Per-layer memory states make training expensive to store and move, so the implementation reconstructs them during backpropagation and batches writes into windows.