Numerical climate models forecast the weather well beyond the deterministic limit. In this “macroweather” regime, they are random number generators. Stochastic models can have more realistic noises and can be forced to converge to the real-world climate. Existing stochastic models do not exploit the very long atmospheric and oceanic memories. With skill up to decades, our new ScaLIng Macroweather Model (SLIMM) exploits this to make forecasts more accurate than GCMs.