We build a nearly-perfect predictive model for memorization in diffusion models using theory and controlled experiments.