The data from three previously published studies on the memory sources of dreams, representing nine different moments of awakening throughout the night, are re-examined. In the original studies, elicited reports were recorded and segmented online into thematic units. The segmented reports were played back to Ss who were asked to identify memory sources or to associate to each segment. Memory sources were classified as episodic, semantic, or abstract self-references. In the meta-analysis and re-analyses reported here, the mean percentages of episodic memory sources are plotted separately for NREM and REM awakenings throughout the night. Within stages, neither NREM nor REM mean percentages differ significantly from each other, whereas between stages the mean percentage of episodic memory sources is significantly greater for NREM than for REM. Even when the correlation between report length and sleep stage is controlled by computing memory source density, the stage effect throughout the night persists for episodic memory sources. The relatively flat episodic memory curves for both NREM and REM indicate a rather constant recruitment of episodic memory sources throughout the night. No stage effect was found for strictly semantic memory sources. When semantic memory was defined generically, however, to include all non-spatio-temporal, "unmarked," information of self as well as of world, significantly more generic semantic memory sources derived from REM than from NREM reports, though not when corrected for the length of dream reports.