A bad evening is supposed to feel less bad by morning. The details should survive; the sting should fade. A new preprint from the University of Hong Kong offers a fine-grained account of how the sleeping brain manages that trick. It does not happen through one uniform process, but through two distinct microscopic events, each confined to a different sleep stage. During deep, dreamless NREM sleep, coordinated brain rhythms appear to anchor the memory in time. During REM sleep, brief bursts of rapid eye movements open a window in which the emotion attached to the memory is drained away.
The study, posted to the bioRxiv preprint server on August 10, 2026, used a within-subject crossover design. Forty-three healthy adults (34 women; average age 24.4) spent two nights in a sleep laboratory one week apart, wearing high-density 64-channel EEG caps while they slept. On one night they watched neutral film clips; on the other, trauma-analog clips of car wrecks, plane crashes, and shootings. The order was counterbalanced, and the two clip sets were matched for semantic content, so that only emotional intensity differed. Thirty-three participants also contributed usable overnight EEG.
To know whether these experiences were replayed during sleep, the team needed a way to see emotional content in brain activity. They first trained a machine-learning classifier on daytime EEG from a functional localizer in which people looked at aversive and neutral images. The classifier learned to distinguish the neural signature of aversive processing from that of neutral processing. It was then applied to the overnight recordings, a multivariate pattern analysis approach. Whenever sleeping brain activity briefly resembled the wakeful signature of emotional processing, the researchers counted that moment as reactivation of the emotional memory.
Reactivation did not spread evenly across the night; it surfaced in exactly two kinds of microevents and no others.
The NREM window: a coordinated duet
Slow-wave sleep is built from two famous rhythms. Cortical slow oscillations are deep, slow waves around 0.5 to 1.25 Hz; sleep spindles are 11 to 16 Hz bursts generated in the thalamus. Decades of research suggest these rhythms matter most when they are coordinated. A spindle that rides the rising phase of a slow oscillation, forming an SO-spindle complex, is thought to mark the moment the hippocampus hands a memory to the neocortex for long-term storage. In this study, valence-specific decoding rose above chance only around these coupled events, roughly 700 to 850 milliseconds after the slow oscillation trough, exactly where the spindle emerges. Isolated slow oscillations carried no decodable signal, and isolated spindles carried none either. The coordinated pair, and only the coordinated pair, carried the emotional content.
The coupling also predicted behavior. Participants estimated the temporal distance between pairs of still frames taken from the films, a measure of how well they remembered when things happened. After sleep, the temporal-distance error for the neutral clips dropped significantly (t = 3.57, p < .001), while aversive events, already remembered more precisely before sleep, remained stable. And people with more frequent SO-spindle coupling showed smaller post-sleep errors (F = 4.66, p = .037). The NREM window, in other words, does the work of an archivist: it files the sequence and preserves the timeline.
The REM window: bursts that soothe
REM sleep is not a uniform state either. It alternates between quiet tonic periods and phasic episodes punctuated by clusters of rapid eye movements. Phasic REM is a distinctive microstate, marked by transient surges of cortical, limbic, and autonomic activation, and the eye movements themselves are time-locked to activity in the amygdala and hippocampus. The preprint shows this is where emotional reactivation lives. Decoding rose above chance between 400 and 1500 milliseconds after the peak of a rapid eye movement, and stayed at chance during matched tonic control periods.
Because phasic REM is defined by eye movements, the obvious worry is that the decoder was reading ocular noise rather than brain activity. Two control analyses addressed this. After removing eye-movement-dominated components identified by independent component analysis, the decoding persisted. Decoding based on those ocular components alone stayed at chance. The signal was cortical, not ocular.
The strength of phasic REM reactivation predicted how much emotional load lifted overnight. Stronger reactivation accompanied sharper drops in self-reported anxiety (r = -.46, p = .007) and negative affect (r = -.44, p = .011), and larger increases in positive affect (r = .47, p = .006). The behavior told the same story: by morning, the gap between the two sessions in negative and positive affect had closed, and the aversive films were rated less negative and less traumatic than they had been the night before. State anxiety, however, remained somewhat elevated.
Two jobs, one night
The two windows did more than coexist; they showed opposite relationships with memory. More phasic REM went with worse temporal-distance memory (F = 6.49, p = .014), the mirror image of the SO-spindle benefit. The windows were also separated in time across the night: SO-spindle reactivation concentrated in the second 90-minute segment of sleep, phasic REM reactivation in the fourth, echoing the natural architecture of deep sleep early and REM later.
Why it matters
For clinicians, the appeal is precision. Sleep-based interventions for conditions such as PTSD already try to reshape memories during sleep, often by cueing them with sounds. Knowing that NREM reactivation rides on SO-spindle complexes while emotional dissipation rides on phasic REM suggests more specific targets: strengthen the coupled events to consolidate structure, or engage the phasic windows to soften affect. The findings also sharpen the long-standing idea that REM sleep acts as a form of overnight therapy by pointing to the exact sub-second moments within REM where the therapeutic work appears to happen.
Limits
The authors flag the study’s limits. The design is correlational: reactivation was measured, not manipulated, so causal claims remain out of reach. The decoder was trained to distinguish emotional valence, not to read memory content, so it cannot show whether the replay preserved the order in which events unfolded during sleep. Scalp EEG cannot localize the neural sources of the effect; that would require intracranial recordings or sleep fMRI. And the trauma-film paradigm, however realistic, is an analog of trauma rather than trauma itself, leaving open whether these mechanisms hold in clinical populations. The work is also a preprint, not yet peer-reviewed.
Bottom line
Sleep, the preprint argues, is not one process with one job. It is a two-act night. NREM couples its slow oscillations and spindles to fix an experience in time, and phasic REM uses its bursts of eye movements as brief windows in which the emotional charge of that same experience is allowed to drain. Remember the timeline, release the feeling. That, the data suggest, is what a good night’s sleep is for.
Source
Zhang, Y., Yao, Z., Chen, D., Xia, T., Zhang, L., Luo, A. F., & Hu, X. (2026). Dissociable reactivation during NREM and REM sleep supports memory consolidation and emotional dissipation. bioRxiv preprint, posted August 10, 2026. DOI: 10.64898/2026.08.04.742295.