In this paper, we propose a framework appropriate for any MPNN architecture that distributes information based on the curvature of the graph's edges. Experiments conducted on different datasets show that our method mitigates over-squashing and outperforms existing graph rewiring methods in several node classification datasets.