Researchers and practitioners employ a variety of time-series processes to forecast betas, either using short-memory models or implicitly imposing infinite memory. We find that both approaches are inadequate: betas show consistent long-memory properties. For the vast majority of stocks, we reject both the short-memory and difference-stationary (random walk) alternatives. A pure long-memory model reliably provides superior beta forecasts compared to all alternatives. Accounting for long memory in beta also pays off economically for portfolio formation. We widely document the robustness of these results.