What "replay" actually refers to
Replay is the reactivation of the same patterns of brain activity that occurred while an animal or person was learning something. In rodents, researchers record hippocampal cells while the animal explores a maze, then look for the same sequences firing again during sleep. Finding them is the operational definition of replay, and it is a measurement of neural activity, not of dreaming or of remembered content.
The strongest replay evidence has historically come from NREM sleep, which is why NREM is often described as the stage where recent memories are replayed [3]. A computational model of slow-wave sleep builds consolidation around exactly this replay step, training an episodic network and a semantic network under sleep-like conditions and then showing that memories can later be retrieved without the hippocampus [3]. That is a model result, not a direct recording, but it reflects the assumption that NREM replay is the well-established case.
REM sleep is not a replay-free zone
A direct test in mice undercuts the simple division of labor. Using hippocampal CA1 recordings combined with Bayesian decoding and sequence analysis, researchers confirmed that representations of awake exploration reappear during subsequent REM episodes, with varying degrees of temporal structure [2]. In other words, REM contains reactivation too, not just NREM.
The same study found that temporally organized sequences were replayed more often during REM sleep after a novel experience, suggesting that salient new experiences sharpen the temporal structure of what gets reactivated in REM [2]. That is a statement about novelty and sequence structure, not about clock time in the night, and it does not establish that REM out-replays early NREM.
Why the early-night comparison is still open
The question mixes two things: which sleep stage replays recent memories, and when in the night that stage dominates. The studies here address the first but not the second. The REM reactivation work compares REM with nREM and with novel versus familiar experience, but does not examine whether early-night NREM is more replay-prone [2]. The slow-wave model simulates NREM consolidation and then simulates REM conditions, but it does not model the night's timeline [3].
Reviews point in a related but different direction. One review attributes memory reinforcement, meaning the ability to retrieve a memory at all, primarily to non-REM sleep, and memory refinement, meaning the precision of that retrieval against competitors, primarily to REM sleep, based on evidence including bidirectional synaptic plasticity in REM and selective REM deprivation [5]. Another review discusses distinct and synergistic roles for the two stages, including replay of recent and remote memories, but its available summary does not isolate early-night NREM [4]. A theoretical framework assigns slow-wave sleep a memory-compression role and REM a generative replay role, yet explicitly does not address early-night timing [1]. So the stage contrast is supported; the clock-time contrast is not.
What would actually answer the question
To test the claim, you would need recordings across the whole night, with replay events tagged by whether they match recent learning, and enough early-night NREM and early-night REM sampled to compare rates fairly. None of the studies described here report that design. The closest supporting logic is indirect: if reinforcement of recent memories is mainly a non-REM job [5], and non-REM's slow-wave period is concentrated early in the night, early NREM would be the expected place to look. That is an inference from stage function plus sleep architecture, not a measured comparison.
The practical takeaway is that "early NREM replays recent memories more than REM" is a reasonable hypothesis with partial grounding, not an established finding. What is established is that NREM replay is well supported [3], that REM also reactivates awake experience and does so more structurally after novelty [2], and that the two stages appear to contribute different things to memory, with reinforcement leaning non-REM and refinement leaning REM [5].
Sources used in this answer
Dreaming as Model Debugging: REM/NREM Sleep as Adversarial Training, Replay, and Precision Reset for Predictive Brains
Proposes a theoretical framework in which slow-wave sleep supports memory compression and REM supports generative replay, but does not address early-night timing or early NREM replay.
Structured and unstructured reactivations during REM sleep are modulated by novel experiences
Confirms that awake exploration is reactivated in hippocampal CA1 during REM sleep, with temporally organized sequences replayed more often after a novel experience.
Consolidation of memory during Slow Wave Sleep and implications in Rapid Eye Movement Sleep: a possible Neural Mass Model
Presents a computational model in which NREM slow-wave replay drives corticalization of episodic memories, while simulated REM conditions produce fragmented, recombined patterns.
Sleep EEG oscillations and synaptic plasticity in memory consolidation across NREM and REM sleep
Reviews distinct and synergistic roles of NREM and REM sleep in memory consolidation, referencing replay of recent and remote memories without isolating early-night NREM.
Contributions of post-learning REM and NREM sleep to memory retrieval
Reviews post-learning sleep effects and attributes memory reinforcement primarily to non-REM sleep and memory refinement primarily to REM sleep.