Remembering information from continuous past episodes is a complex task. On the one hand, we must be able to recall events in a highly accurate way that often includes exact timing; on the other hand, we can ignore irrelevant details and skip to events of interest. We here track continuous episodes, consisting of different sub-events, as they are recalled from memory. In behavioral and MEG data, we show that memory replay is temporally compressed and proceeds in a forward direction. Neural replay is characterized by the reinstatement of temporal patterns from encoding. These fragments of activity reappear on a compressed timescale. Herein, the replay of sub-events takes longer than the transition from one sub-event to another. This identifies episodic memory replay as a dynamic process in which participants replay fragments of fine-grained temporal patterns and are able to skip flexibly across sub-events.