Memory Reactivation during Sleep Does Not Act Holistically on Object Memory

Abstract

Memory reactivation during sleep is thought to facilitate memory consolidation. Most sleep reactivation research has examined how reactivation of specific facts, objects, and associations benefits their overall retention. However, our memories are not unitary, and not all features of a memory persist in tandem over time. Instead, our memories are transformed, with some features strengthened and others weakened. Does sleep reactivation drive memory transformation? We leveraged the Targeted Memory Reactivation technique in an object category learning paradigm to examine this question. Participants (20 female, 14 male) learned three categories of novel objects, where each object had unique, distinguishing features as well as features shared with other members of its category. We used a real-time EEG protocol to cue the reactivation of these objects during sleep at moments optimized to generate reactivation events. We found that reactivation improved memory for distinguishing features while worsening memory for shared features, suggesting a differentiation process. The results indicate that sleep reactivation does not act holistically on object memories, instead supporting a transformation where some features are enhanced over others.

Keywords: NREM sleep; category learning; consolidation; replay; targeted memory reactivation.

Figure 1.

Experimental design. A, Study timeline and stimuli. Participants completed a novel object category learning task and took a nap where TMR was performed. They studied three categories of novel “satellite” objects (alpha, beta, gamma), where a satellite could have parts that were unique to itself and parts shared with other members of its category. Each satellite had its own unique name (e.g., nivex). B, Learning phase. First, participants were exposed to the satellites one-by-one: participants heard a satellite's name out loud and then saw it on screen. Next, participants completed blocks of trials where they first heard a satellite's name and then were shown a satellite with one feature missing and instructed to select the missing feature. C, Test phase. Participants heard a satellite's name and then were shown a satellite with one or two features missing. They selected a feature and then rated their confidence in their decision.

Figure 2.

Pre- and post-nap behavior. A, Overall accuracy across learning blocks. B, Pre-nap and post-nap test performance. Mean accuracy on shared (purple) and unique (orange) feature memory trials. The dotted line indicates chance. C, Assessment of tradeoff in unique and shared feature accuracy. A line was fit, for each participant, predicting shared feature accuracy from unique feature accuracy of the corresponding object, for the pre-nap (left) and post-nap (middle) tests. Mean slope across participants is represented with a thick dotted line. A negative slope indicates a tradeoff in shared and unique feature accuracy. Right, Barplot of slopes in the pre-nap (blue) and post-nap (red) tests. Bars represent mean slopes across participants. Error bars represent ±1 SEM. Dots and lines correspond to individual participants. *p < 0.05, **p < 0.01, ***p < 0.001.

Figure 3.

Real-time TMR protocol. A, Overview of TMR. (left) TMR cues—delivered aloud as individual satellite names—began after a participant entered and remained in N2 sleep for at least 3 min. One category was cued in interleaved order, one in blocked order, and one was left uncued. The presentation of the two cued categories was intermixed across NREM sleep. See Extended Data Figure 3-1 for more details. B, Real-time TMR administration. We developed a novel TMR protocol in which cues were played in the up-states of slow oscillations (up-state = signal goes above threshold of +35 μV; Ngo and Staresina, 2022) and at least 2.5 s after a detected spindle (Antony et al., 2018b). C, Left, TFR of the difference between the neural response when a TMR sound cue was played versus not (sham), with the ERP response to sound cues overlaid (averaged across all channels and all participants). Sounds were played at time = 0. Right, Cluster-based permutation testing identified two significant clusters in the time-frequency response to TMR sound cues. Unshaded areas represent clusters identified via performing t tests across participants (α = 0.01). Clusters that survived subsequent permutation testing are highlighted in white. Topoplots show the spatial representation of the identified clusters. See Extended Data Figure 3-2 for more details.

Figure 4.

TMR cueing improved unique feature memory and impaired shared feature memory. A, Impacts of TMR. A linear mixed-effects model was used to analyze change in unique and shared feature accuracy across the nap as a function of cueing. Model estimates for accuracy change for shared (purple) and unique (orange) features are plotted. B, Replication of analysis from A using only the subset of items whose shared feature accuracy was greater than or equal to their unique feature accuracy. Left, Mean shared and unique feature accuracy from the pre-nap test. Dots represent participants, error bars represent ±1 SEM, and the dotted gray line represents chance. Right, Model estimates of unique and shared feature accuracy change for uncued and cued items from this subset. C, Impact of cueing on uncued items from cued categories. Plotted are the model estimates for unique and shared feature accuracy change for items in the designated “uncued” category (uncued–uncued), items uncued in a designated “cued” category (uncued–cued), and for items that were cued (cued–cued). *p < 0.05, **p < 0.01, ***p < 0.001. See Extended Data Figure 4-1 for additional model details.

Figure 5.

Blocked cue presentation led to greater memory change than interleaved. A, Model estimates of accuracy change for items that were cued in interleaved or blocked order. B, Model estimates of accuracy change for fully cued blocked items, as a function of their sequence in the cueing order. Sequence position 4 corresponds to the last item cued before the participant woke up from their nap; 0 corresponds to items in the uncued category. Plotted are model estimates for accuracy change for unique and shared features, with a linear fit to sequence position (shaded area = 95% confidence intervals).

Figure 6.

Cueing as a driver of differentiation. A, Schematic of representational overlap for four exemplars from the same category. B, If cued objects differentiate, the unique aspects of their representations are enhanced and the shared aspects are diminished. These effects generalize to an uncued item, because the differentiation of the other exemplars reduces overlap for all items, and all unique elements are subject to less interference.