This study investigates the realized volatility of major cryptocurrencies with a focus on long memory properties. Using high-frequency data, the study evaluates whether long-memory properties improve volatility modelling and forecasting performance. To this end, the four well-known time series models are compared: the ARFIMA, which explicitly captures fractional integration and long memory; the HAR model, which approximates long-memory behaviour through multi-horizon components; and the benchmark AR model and ARMA model specifications. The results reveal strong persistence in cryptocurrency volatility consistent with long-memory behaviour. However, the HAR model consistently outperforms alternatives in both in-sample and out-of-sample forecasts, suggesting that multi-scale structures capture long-memory dynamics more effectively than the ARFIMA model. Overall, the findings highlight that while long memory is a key feature of cryptocurrency volatility, it is better approximated through heterogeneous temporal components, with important implications for forecasting and risk management.