This paper focuses on the implications of asset return predictability on long-term portfolio choice when return forecasting variables exhibit long memory. Recent research in empirical finance argues that expected asset returns are time-varying and relates them to various predicting variables that historically reveal very gradual movements in time; hence, we aim at careful modelling of their persistence properties. For that purpose, we exploit the class of fractionally integrated processes. Our theoretical derivations indicate profound impact of the long-memory component on optimal long-term portfolio weights. We illustrate our approach to the modelling of asset return dynamics on post-war US data for equities, Treasury bonds, and cash.