Multiple memories can be simultaneously reactivated during sleep as effectively as a single memory

The study investigates whether multiple memories can be reactivated during sleep without mutual interference, addressing whether reactivation resources are limited or can operate in parallel. Building on the active systems consolidation framework, which posits hippocampal-cortical reactivation during sleep, prior work has demonstrated replay in animals and reactivation in humans using fMRI/MEG and targeted memory reactivation (TMR) via cues (odors or sounds). However, it remains unclear if reactivation benefits scale with the number of items associated with a cue. The authors contrast a limited capacity hypothesis (LCH)—predicting diminished per-item benefits as set size increases—against a parallel reactivation hypothesis (PRH)—predicting set-size-independent per-item benefits. The research tests these models by associating sounds with sets of 1, 2, or 6 semantically related items learned in a spatial-memory task and re-presenting sounds during NREM sleep.

The paper reviews evidence for memory reactivation during sleep from rodent replay studies and human neuroimaging showing reactivation during sleep and rest. TMR studies using odors (often cuing many items with one odor) and sounds (often one or two items per sound) reliably enhance memory, as supported by a meta-analysis. Prior work has suggested capacity limits for sleep-dependent consolidation and competitive interactions, but no direct comparison had been made of single-item versus multi-item reactivation benefits using the same cue modality within-subjects. The authors also consider models where only a subset of items is reactivated (randomly or biased toward certain items), which predict set-size-dependent benefits and/or relationships with cue repetition count.

Participants (inferred N≈31 from ANOVA df) learned locations of 90 images on a 2D circular grid. Images were organized into semantically coherent sets of size 1, 2, or 6; each set was associated with a congruent auditory cue (e.g., meow, ring). Learning employed feedback-based placement trials to criterion. Pre-sleep spatial memory was tested without feedback. During a 90-minute afternoon nap, EEG (64-channel BioSemi ActiveTwo; 512 Hz) and polysomnography were recorded. Auditory cues associated with half of the sets were presented unobtrusively during NREM (N2/N3) sleep (on average ~11 repetitions per sound; ~10 during N2/N3). A novel sound was also presented for physiological control. Post-sleep memory was tested identically to pre-sleep. To handle potential confusions among similar items within a set, the authors identified and excluded swap errors (misplacing an item near another item’s location) from accuracy analyses. Behavioral analyses included repeated-measures ANOVAs on pre/post error rates, cuing status (cued vs non-cued), and set size, as well as complementary analyses to address pre-sleep differences (regression adjustments and subsampling/bootstrapping). They further assessed within-set relationships using intraclass correlation (ICC) and outlier-like Z-score analyses to test subset-reactivation models. EEG analyses computed time-frequency responses (STFT) time-locked to cue onset, with cluster-based detection of cue-modulated power, focusing on delta-theta (approx. 0.5–8 Hz) and sigma (11–17 Hz) bands. Spindles were automatically detected to estimate time-locked spindle probability and amplitude; linear mixed models assessed modulation by set size and cue repetition count, and related physiology to behavioral cueing benefits.

- Learning/training: Number of trials to criterion did not differ by cuing status or set size; no interaction (e.g., F(1,30)=0.66, p=0.42; F(2,60)=1.21, p=0.30). - Pre-sleep: Error rates did not differ between cued and non-cued sets (F(1,30)=0.74, p=0.27); set size differed (F(2,60)=7.52, p<0.001) with smaller errors for single-item sets; no interaction (F(2,60)=0.14, p=0.87). Swap errors were more frequent in larger sets but did not differ by cuing status pre-sleep. - Cueing benefit per item: Significant cuing effect with better recall for cued vs non-cued items (e.g., F≈13.4, p<0.001; η²≈0.31). No main effect of set size (F≈0.60, p=0.52) and no cuing×set-size interaction (F≈0.42, p=0.66). Bayes factor indicated strong evidence for no interaction (BF10=5.25 favoring null interaction). On average, spatial error decreased by 2.98±1.87 pixels for cued sets and increased by 4.76±2.39 pixels for non-cued sets. - Robustness: Complementary analyses (regression adjustment; 500 subsampled datasets controlling pre-sleep differences) converged: significant cuing effects in most subsamples (67% p<0.05) and virtually no cuing×set-size interactions (0.6%). - Cumulative per-set benefit: Significant cuing effect (F(1,30)=98.3, p<0.001), no main effect of set size; a marginal interaction suggested larger cumulative benefits for larger sets due to more items per set (F(2,60)=3.02, p=0.056). - Subset-reactivation models: Within-set ICC and outlier analyses did not support biased subset reactivation; no pattern consistent with a small subset driving benefits. Random subset model predictions were not supported: no set-size-dependent per-item benefits and no correlations between cue repetitions and cuing benefit (e.g., r=0.09, p=0.63 for single items; r=0.01, p=0.95 for 2-item sets; r=0.29, p=0.69 for 6-item sets). Probability-based estimates also showed no correlation (r’s ~0.13–0.19, ns). - Physiology: Cue presentations during NREM increased power in delta-theta and sigma bands. Both bands showed linear modulation by set size—larger sets produced greater power increases (delta-theta: t(6889)=2.52, p<0.02; sigma: t(6889)=2.88, p<0.004). Spindle probability (time-locked to cue) increased with set size (t(6889)=2.22, p<0.03), whereas spindle amplitude did not (t(1608)=0.44, p=0.65). These physiological effects were not driven by sound repetition counts. Physiological measures did not significantly predict per-item cueing benefits. Sleep time measures did not correlate with cueing benefit.

Findings demonstrate that targeted memory reactivation during sleep benefits cued items regardless of whether cues are linked to one, two, or six items, supporting a parallel reactivation hypothesis rather than a limited-capacity model. Analyses failed to support models where only a subset of items is reactivated (randomly or biased), as there was no set-size dependence of per-item benefits and no relationship with the number of cue repetitions. EEG results showed that delta-theta and sigma power, as well as spindle probability, scale with the number of items associated with the cue, indicating that the extent of prior learning modulates cue-evoked neural responses during sleep. The authors consider whether rapid sequential reactivation could explain the behavioral results but argue that the observed physiology and time courses are more consistent with parallel reactivation. They also propose an alternative model in which cues reactivate a generalized context that in turn benefits all embedded items, aligning with context-based memory theories (e.g., CMR). This context reactivation could explain sustained effects after cue offset and scaling of physiological responses with set size. The apparent discrepancy between set-size-independent per-item benefits and set-size-dependent physiological responses is reconciled by noting that cumulative set-level benefits increase with set size, while per-item benefits remain constant, suggesting physiology may index the scope of reactivation rather than per-item gain.

The study shows that multiple memories associated with a single cue can be reactivated during sleep as effectively per item as a single memory, supporting a parallel, not strictly limited, capacity for sleep-related consolidation. Cue-evoked delta-theta and sigma activity, including increased spindle probability, scale with the number of associated items, indicating that neural responses reflect the extent of learned associations. The work challenges limited-capacity reactivation models and raises the possibility that sleep cues can reactivate generalized contexts that benefit embedded items. Future research should delineate boundary conditions and capacity limits (e.g., larger set sizes), test the role of semantic relatedness and context, examine individual differences in simultaneous reactivation capacity, and further disambiguate parallel from fast sequential reactivation using temporally resolved neural measures.

- Sample mainly comprised university students, limiting generalizability and leaving open the role of individual differences (e.g., working memory capacity) in simultaneous reactivation. - Stimulus categories were highly related and congruent with sounds, potentially encouraging chunking or context-level representations; results may not generalize to unrelated items or other modalities. - Variability in relatedness across sets could introduce noise; although balanced algorithmically, residual differences may affect outcomes. - Design choices (e.g., set sizes limited to 1, 2, 6; spatial task specifics) constrain inference about broader capacity limits; sets larger than six may reveal interference. - Differences across TMR paradigms in the literature (e.g., pairwise spatial TMR reporting anti-correlated benefits) complicate direct comparisons and may indicate boundary conditions where subset reactivation occurs.