In this paper we remark that the evolution of the realized volatility is characterized by a combination between high–frequency dynamics and a smoother persistent dynamics evolving at a lower–frequency. We suggest a new Multiplicative Error Model which combines the mixed frequency features of a MIDAS with Markovian dynamics. When estimated in–sample on the realized kernel volatility of the S&P500 index, this model dominates other simpler specifications, especially when monthly aggregated realized volatility is used. The same pattern is confirmed in the out–of–sample forecasting performance which suggests that adding an abrupt change in the average level of volatility better helps in tracking extreme episodes of volatility and a relative quick absorption of the shocks.