Dreaming of a Learning Task is Associated with Enhanced Memory Consolidation: Replication in an Overnight Sleep Study

Summary

Sleep following learning benefits memory. One model attributes this effect to the iterative “reactivation” of memory traces in the sleeping brain, demonstrated in animal models. Although technical limitations prohibit using the same methods to observe memory reactivation in the human brain, the study of mental activity during sleep provides an alternative method of observing memory activation during sleep. In fact, the content of dream experience may reflect the process of memory reactivation and consolidation in the sleeping brain. In line with this hypothesis, we previously reported that dreaming about a spatial learning task during a nap strongly predicts subsequent performance improvements. Here, we replicate this observation in an overnight sleep study, for the first time demonstrating that pre-sleep training on a virtual maze navigation task is reflected in dreams reported from all phases of sleep, with unambiguous representation of the task in dream content associated with improved next-morning performance. These observations are consistent with reactivation-based models of memory consolidation in sleep, confirming our earlier finding that the cognitive-level activation of recent experience during sleep is associated with subsequent performance gains.

Introduction

Post-training sleep is beneficial for human memory (Diekelmann and Born, 2010; Stickgold, 2005). Neurophysiological studies in rodents suggest that this is due to an iterative “reactivation” of memory networks during sleep, strengthening synaptic connections and facilitating memory consolidation (Ji and Wilson, 2006; Wilson and McNaughton, 1994). Human studies have also reported analogous observations, for example in PET studies finding that patterns of neural activity present during encoding are reiterated during subsequent sleep (Maquet et al., 2000; Peigneux et al., 2004). During sleep, the activity of memory networks is also evident in the content of subjective experience, as recent waking experiences are often unambiguously represented in the content of our nightly dreams (Stickgold et al., 2000; Wamsley and Stickgold, 2009). Indeed, the observation that events from the previous day appear in dream content stretches back more than a hundred years (Freud, 1913). Yet only recently has the effect of waking experience on dream reports been studied in the laboratory using scientifically rigorous methods. Recently, we reported that dreaming of a learning task was associated with improved memory for that information following sleep (Wamsley et al., 2010b). This observation echoes earlier work in non-laboratory contexts, for example showing that participants enrolled in a language immersion course show superior grade outcomes when their dreams incorporate the foreign language (De Koninck et al., 1990).

The observation that patterns of neural activity representing recent experience are “reactivated” in the sleeping brains of rodents (e.g. Wilson and McNaughton, 1994) suggests that sleep’s effect on human memory might be attributed to a repeated “replay” of experience during sleep allowing cortical memory networks to be strengthened and integrated. Indeed, recent studies demonstrate that hippocampal sharp-wave ripples (during which memory reactivation preferentially occurs) increase following learning (Eschenko et al., 2008), and that experimentally disrupting ripples impairs memory (Ego-Stengel and Wilson, 2010). Still, alternative accounts attribute the memory benefits of sleep to other mechanisms. For example, a recent extension of the “synaptic homeostasis hypothesis” suggests that a selective downscaling of synaptic strength during sleep could facilitate memory by enhancing signal-to-noise ratio, without the need for invoking either memory reactivation or synaptic potentiation (Nere et al., 2013). In fact, despite the evidence described above, there has not yet been a direct demonstration that the type of cellular-level memory reactivation observed in rodents even correlates with sleep’s beneficial effect on memory in humans.

Demonstrating the reactivation of specific memory networks during human sleep has been challenging due to technical limitations – the large scale single-cell recordings enabling this work in rodents are not available in human subjects, and the spatial and temporal resolution of noninvasive neuroimaging methods are inadequate to the task. Our own work has focused on mental activity during sleep as information about the activation of memory networks. Our laboratory and others have demonstrated that memories formed during engaging, interactive visuospatial learning tasks are clearly incorporated into dream experience during post-training sleep (Kusse et al., 2011; Solomonova et al., 2015; Stickgold et al., 2000; Wamsley et al., 2010a). These dream reports are not only “symbolically” or weakly related to prior experience. For example, after playing the video game “Tetris”, 63% of participants reported unambiguous imagery of the game during the first minutes of sleep, with one participant reporting:

…just seeing Tetris shapes floating around in my head like they would in the game, falling down, sort of putting them together in my mind

(Stickgold et al., 2000).

In another study, after playing the downhill skiing arcade game Alpine Racer II, 23% of reports contained unambiguous reference to the game, at times including imagery of specific locations on the virtual slope. For example, one participant in these studies reported that:

I get like flashes of that … game in my head, virtual reality skiing game … downhill umm race, in my head. Umm, there’s this one particular corner that I haven’t quite been able to master, and every time I get flashes of it, it’s like that corner that umm I keep crashing into

(Wamsley et al., 2010a).

Although these reports were collected from the first few minutes of sleep (the “sleep onset” period), we later found incorporation of a virtual maze navigation task during N2 sleep further into the sleep cycle of a daytime nap (Wamsley et al., 2010b). Together with other evidence (De Koninck et al., 1990; Kusse et al., 2011; Nielsen and Stenstrom, 2005; Tauber et al., 1968), these observations provide a clear demonstration that memory networks encoding recent learning experiences are activated during sleep.

This evidence raises the question of whether the activation of memory networks underlying dream content is related to sleep-dependent memory consolidation and to the effects of sleep on human memory. Importantly, this is not necessarily the case—the appearance of recent experience in dreams could be driven by neural processes entirely independent of those which support the consolidation of memory. However, we have previously reported that participants who dreamed about a virtual maze navigation task during a daytime nap showed a tenfold greater performance gain at subsequent test than those whose dream reports were unrelated to the task, supporting this hypothesis (Wamsley et al., 2010b). This observation suggests that dreaming of recent experience reflects the reactivation and consolidation of memory in the sleeping brain, which in turn leads to post-sleep performance improvements.

But rigorous science demands the replication of surprising new observations before accepting them as fact. The generalizability of our original study could be criticized on several grounds. First, these observations were based on only a small number of participants who incorporated the maze task into dream content, and dreams were collected only from NREM sleep during a short daytime nap. Second, several subsequent studies testing the same hypothesis using either a different learning task (Schredl and Erlacher, 2010), or a modified version of the VMT (Stamm et al., 2014; Wamsley et al., 2016), failed to find any association between incorporation of the task into dreaming and subsequent memory.

In the current study, we sought (1) to replicate our prior observation that dreaming of the VMT learning task is associated with memory enhancement, (2) to extend this observation to reports from all sleep stages across a full night of sleep.

Methods

Participants

N=39 healthy college students (20±2SD yrs; 33% male) completed the study. Participants reported a mean habitual sleep time of 7.0±1.1SD hrs, and 66.7% of the sample reported habitual napping on an average of 3.6±1.6SD days/week. No participants reported diagnosis of a sleep disorder or mental disorder. One participant was under treatment with an antibiotic, and no others reported current medication use. Frequency of 3-D style video game play averaged 2.1 on a 5-point ordinal scale, with the majority of participants reporting game play at least once per month (35%) or once per year (30%), with a minority of participants reporting game play once a week (22.5%) or every day (5%), or less than once a year (7.5%). Participants were randomly assigned to either a “Sleep” group (n=21) or an “Immediate Test” control group (n=18). Two additional control participants enrolled in the study but failed to complete testing due to “cybersickness” induced by the Virtual Maze Task (VMT).

Following criteria established in our earlier work (Wamsley et al., 2010c), participants were excluded from VMT performance analyses if they reported being excessively sleepy at the time of maze training (Stanford Sleepiness Scale (SSS) ≥5) or playing 3D style video games less than once per year (Wamsley et al., 2010c). Following exclusions, n=17 Sleep and n=16 Immediate Test participants were included in analyses of VMT performance. Based on the very large effect size reported in our previous paper (d=2.02 (Wamsley et al., 2010b)), this sample size was adequate to test our primary hypothesis with power >0.9. This research was approved by Beth Israel Deaconess Medical Center’s IRB.

Procedures

Procedures closely followed those in our prior study of dreaming during daytime naps (Wamsley et al., 2010b), with the primary exception being that here we tested participants across a full night of sleep (Fig 1). Participants in the Sleep group reported to the laboratory at approximately 9PM, signed written informed consent, and completed questionnaires including a demographics information form, 3-day retrospective sleep log, Epworth Sleepiness Scale (ESS), and Beck’s Anxiety Inventory. Participants were then wired for polysomnographic recording, with 58 EEG electrodes placed according to the international 10–10 system of electrode placement, bipolar chin electrodes for recording electromyography, and electrooculography electrodes placed at the left and right outer canthus. At ≅11PM, participants completed the SSS as a measure of state sleepiness before reclining with their eyes closed for a 10min baseline EEG recording of resting wakefulness. Participants were then trained on the Virtual Maze Navigation Task (VMT, described below) before undergoing a second 10min resting EEG recording and completing visual analog scales describing subjective task engagement and difficulty. Lights were then turned off and participants began an 8.5hr sleep opportunity, during which subjective experience was sampled as described below. Sleep participants were awoken at 8am the following morning and were retested on their ability to solve the same maze approximately 30min later.

Figure 1.

Experimental timeline.

Sleep participants were prompted to provide open-ended verbal reports on their mental experience during both the waking rest periods and nocturnal sleep. During each 10min rest period, participants were prompted to provide 2 reports. During the first hour of sleep, participants were awakened to provide up to 10 sleep onset reports, after 30sec, 60sec, or 90sec of PSG-defined sleep (order of sleep latency awakenings pseudorandomized). During the remainder of the night, participants were awoken three times from N2 NREM sleep (on each occasion following at least 10min of continuous N2) and provided one final report upon morning awakening, regardless of sleep stage. Subsequent analysis showed that, amongst the full sample of participants, morning awakenings were from N2 sleep in 10 participants, and were from REM sleep in 10 participants. For one participant, sleep stage upon morning awakening could not be determined due to excessive artifact in the EEG record. For each report, participants were contacted by calling their name over an intercom, at which point they were prompted to provide an open-ended verbal report describing “everything that was going through your mind” just prior to awakening.

Participants in the Immediate Test control group trained and were tested on the maze without intervening sleep. Immediate Test subjects reported to the laboratory and completed the same initial paperwork as Sleep group participants prior to training on the VMT, as described as below. Following training, participants took a 15min break during which they were allowed to stand up and move around the laboratory, after which they were immediately tested on the maze. No mentation reports were collected from Immediate Test participants.

Virtual Maze Task

The Virtual Maze Learning Task (VMT) has been described in several of our prior papers (Nguyen et al., 2013; Stamm et al., 2014; Wamsley et al., 2010b; Wamsley et al., 2010c). Constructed using the Unreal Tournament 3 Editor (Epic Games), the task is implemented on a PC computer and projected onto a screen in a darkened testing room. Participants navigate through the virtual environment using a USB number pad with four arrow keys. The training session begins with a 5min exploration period during which participants are placed at the maze exit and instructed to explore as much of the maze as possible and to remember it as well as possible, as they will subsequently be tested on their ability to locate the exit. This exploration period is followed by a series of three training trials in which participants attempt to find the exit as quickly as possible. For each trial, participants began at a different starting point, each equidistant from the exit, with the order of starting positions counterbalanced across participants. In this version of the task, participants were incentivized with a performance-based payment (Stamm et al., 2014). A monetary reward value was displayed on screen at the beginning of each trial, and counted down as the participant played. Participants were told that the remaining money was theirs to keep at the end of the trial. Trials ended when subjects reached the exit of the maze or after a maximum of 10min. Retesting consisted of three additional trials using the same starting-point order as during training, but without the 5min exploration period.

The primary dependent measure of “Completion Time” is the number of seconds required to reach the maze exit on each trial. To facilitate further analysis of movement behavior during the task, the maze space is divided into a 20x20 grid. “Distance Traveled” to the exit is defined as the total number of grid boundaries traversed in each trial. “Backtracking” (1 − [number of unique positions / Distance Traveled]) measures the extent to which participants retrace their steps while navigating towards the maze exit. For all performance measures, overnight improvement was calculated as the change from the last training trial to the mean performance on the 3 retest trials (Wamsley et al., 2010b).

Data Analysis

Report transcripts were scored for content related to the learning task by two raters blind to subject, experimental condition, sleep stage, and maze performance. Following procedures established in our earlier work (Wamsley et al., 2010a), reports were scored first according to whether the participant recalled any mental content, and then according to whether the report directly mentioned the maze task (making explicit reference to the maze task from the experiment or another maze, see Table 1) and/or described content indirectly related to the maze (incorporating sensations, objects, locations, actions or themes similar to those contained in the maze).

Table 1.

Example Maze Task Incorporations

	Task Incorporation Examples
Sleep Onset	“… I can see the maze, and I think if we could swim above it, we could see everything.”
REM Sleep	“… we were just like walking through it but it was like a formal maze … one of those outdoor like ones made out of like hedges and bushes.”
Stage 2 Sleep	“… dreamed I was like standing in the middle of a maze, waiting for my friend to find me. And I was waiting and waiting and she just kept going around and around, calling my name …”
Wake	“I was trying to map out the design of the maze and … I was thinking about, how much money I’m not going to get, which is about ten, fifteen bucks, and that was about it.”

To test whether maze-related content was reported at greater-than-chance levels, raters additionally scored N=100 control dream reports collected in a similar sleep study (Rowley et al., 1998), in which participants were awakened to provide mentation reports without having trained on a learning task prior to sleep. Control reports were randomly intermixed with the experimental reports given to raters for scoring.

Inter-rater agreement was high for the presence of mental content (Cohen’s Kappa=0.75), and for direct incorporation of the learning task into dreaming (Kappa=0.89). Disagreements were resolved by a third rater, also blind to subject, condition, sleep stage, and performance. However, agreement on whether reports indirectly incorporated the learning task was very poor (Kappa=0.07). Scores of indirectly related content were also not correlated with any performance improvement metric. Data on indirect incorporation are therefore not further discussed in this paper.

Results

Incorporation of the Maze Task into Dreaming and Waking Cognition

347 mentation reports were elicited across states of sleep and wake, 310 (89%) of which contained at least some mental content (see Tables 1&2). All told, 39 reports contributed by 12 of the 17 Sleep participants included in analysis (71%) contained content related to the maze task, and these occurred across all conscious states (Table 2). In contrast, no Control reports (Rowley et al., 1998) were judged to contain maze-related content. In all states except for N2, task incorporation was significantly greater than the 0% rate established in the Control set (chi-square tests of independence, Table 2).

Table 2.

Subjective Reports by Conscious State

	N Elicited	N Content	N Task-Related
			#	%	p
Pre-Training Rest	36	36 (100%)	4	11%	0.001*
Post-Training Rest	40	40 (100%)	20	50%	2 x 10−14*
Sleep Onset	188	170 (90%)	12	7%	0.01*
N2 Sleep	62	48 (77%)	1	2%	0.2
Morning N2 and REM‡	21	16 (76%)	2	13%	0.002*
Control Reports	100	64 (64%)	0	0%	-

Notes. P-values derived from chi-square tests comparing task incorporation in each state to the 0% incorporation observed in the control report set.

^*Incorporation of the learning task was significantly greater than in the Control report set.;

^‡50% of Morning reports were elicited upon awakening from N2, and the remaining 50% upon awakening from REM. Of the 2 task-related morning reports, one was recalled from N2, and one was recalled from REM sleep.

Waking rest reports contained incorporation of the learning task even before participants were first exposed to the maze (Table 2). The content of these pre-exposure reports reflected anticipation of the upcoming learning task, which participants had been informed would be occurring shortly. During post-training rest, incorporation of the learning task increased significantly above pre-training levels (52% of reports vs. 11%, Wilcoxon rank sign test, Z=2.60, p=0.009; Fig 2).

Figure 2.

Quiet rest reports. Left: Task-related reports per subject during Pre-Training and Post-Training rest periods. Right: Improvement in completion time (left axis, open circles) and in distance traveled (right axis, solid circles) as a function of the number of task-related reports subjects gave during post-training quiet rest. Error bars: +/− SEM.

Both Dreaming and Waking Cognition Predict Next-Morning Memory Performance

Sleep Reports

Participants that directly incorporated the maze task into at least one dream report (“Incorporators”, n=8) showed greater overnight improvement in both Completion Time (172 vs. 2 sec, t₁₅=2.38, p=0.03, d=1.16; Fig 3, left) and Distance Traveled (150 vs. −2 grid units, t₁₅=2.27, p=0.04, d=1.10; Fig 3, right), than those without direct task incorporations (“Non-Incorporators”, n=9). Incorporators also showed a trend toward decreased Backtracking relative to Non-Incorporators (14% vs. −3%, t₁₅=2.04, p=0.06).

Figure 3.

Task improvement and task-related dream content. Improvement in completion time (left) and distance travelled (right) for subjects who did (“Incorporators”) or did not incorporate (“Non-Incorporators”) the task into at least one dream report. Error bars: +/−SEM.

Waking Rest Reports

Participants whose post-training rest reports directly incorporated the maze task also showed non-significantly greater improvement in Completion Time (112 vs. −17 sec, t₁₄=1.49, p=0.16) and Distance Traveled (71 vs. −17 grid units, t₁₄=1.53, p=0.15), and a trend toward greater improvement in Backtracking (9% vs. −7%, t₁₄=1.83, p=0.09), than those without direct incorporations. The failure of these comparisons to reach significance, however, appears to reflect the difference in improvement between subjects who incorporated the task into one, as opposed to two, reports (Fig 2).

Because incorporation rates were quite high during post-training rest, with a full 50% of post-training rest reports directly related to the maze (Table 2), we were also able to examine how the amount of maze-related content predicted subsequent overnight performance improvements. The number of post-training rest reports related to the maze predicted overnight improvements in navigation performance (Spearman’s rank order correlations: Completion Time ρ=0.50, p=0.051; Distance Traveled ρ=0.53, p=0.04; Backtracking ρ=0.55, p=0.03; Fig 2). The n=3 individuals who incorporated direct references to the maze task into pre-training rest reports did not improve significantly more than other participants (p-values for all three measures >0.17).

VMT Performance

Overall, performance improved substantially across trials (linear polynomial contrast for the decrease in Completion Time across trials: F_1,32=16.45, p=0.0003). Following training, performance change across a night of sleep did not differ significantly from change across 15min in the Immediate Test group (Completion Time: t₃₁=1.62, p=0.11; Distance Traveled: t₃₁=1.25,p=0.22; Backtracking: t₃₁=0.58, p=0.57).

Baseline Performance Predicts Task-Related Dreaming

As in our prior study (Wamsley et al., 2010b), participants who incorporated the learning task into dreaming performed more poorly during training than those who did not. In the current study, “Incorporators” performed more poorly than “Non-Incorporators” on final training trial for Completion Time (439 vs. 212 sec, t₁₅=2.39, p=0.03), Distance Traveled (286 vs. 147 grid units, t₁₅=2.27, p=0.04), and Backtracking (42% vs. 21%, t₁₅=2.10, p=0.05). In contrast, this baseline performance was not correlated with incorporation of the task into post-training rest reports (Completion Time: p=0.59, Distance Traveled: p=0.55, Backtracking: p=0.63). Because Incorporators showed both poor baseline performance and enhanced overnight improvement (which were themselves strongly correlated: r₁₇=0.92, p<0.001), we conducted an ANCOVA testing the effect of task-related dreaming on overnight Completion Time improvement while controlling for baseline Completion Time as a covariate. In this model, the effect of dream incorporation on overnight improvement was no longer statistically significant (p=0.64).

Sleep Architecture

On the experimental night, Incorporators slept an average of 51±19 (SEM) min longer than Non-Incorporators (t₁₄ =2.74 p=0.02), resulting from similar reductions in all sleep stages (Table 3). No significant correlations were found between TST, N2, N3, or REM sleep amounts and post-sleep performance gains (Completion Time: TST: r₁₆=0.18, p=0.51; N2: r₁₆=0.32, p=0.23; SWS: r₁₆=−0.23, p=0.39; REM: r₁₆=−0.24, p=0.37). However, there was a significant correlation between overnight performance improvement and time spent in N1 sleep (Completion Time: r₁₆=0.67, p=0.005), which survives Bonferroni multiple-comparisons correction. We therefore conducted an ANCOVA testing the effect of task-related dreaming on improvement in overnight Completion Time while controlling for the amount of N1 sleep as a covariate. In this model, the effect of dream incorporation on overnight improvement remained statistically significant (F_1,14=6.36, p=0.03).

Table 3.

Sleep and Demographic Characteristics of Incorporators and Non-Incorporators.

	Incorporators		Non-Incorporators		I:NI	p-value
	mean	SD	mean	SD	%
Age (yrs)	19.3	1.4	20.0	1.9	104%	.37
Sex (% male)	25%		22%			.89
Baseline Completion Time	439	244	212	140	48%	.03*
TST (min)	411.1	29.6	360.1	43.5	88%	.02*
N1 (min)	36.5	13.0	26.3	10.5	72%	.10
N2 (min)	228.6	31.5	205.7	28.2	90%	.15
N3 (min)	69.2	26.0	62.1	19.8	90%	.55
REM	76.8	17.1	66.1	23.5	86%	.32
ESS	7.5	4.1	7.4	3.5	99%	.98
SSS	3.0	0.5	2.7	0.5	90%	.20

Notes. Sleep Log TST – mean total sleep time from pre-study sleep log; ESS – Epworth Sleepiness Scale; SSS – Stanford Sleepiness Scale at training; Baseline Completion Time – mean time to solve the maze (sec) in the final training trial. I:NI – mean value for Non- Incorporators as percent of value for Incorporators. p-values are derived from independent samples t-tests, with the exception of sex, which was tested using a Chi-Square test of independence.

Discussion

In support of the hypothesis that dreaming reflects memory processing in the sleeping brain, here we report that participants who incorporate a recent learning experience into their overnight dream content show enhanced overnight performance improvement. Importantly, this observation exactly replicates the findings of our previously published nap study (Wamsley et al., 2010b), using the same criteria for direct task incorporation, the same learning task, and the same methods of assessing performance improvement. We also confirm, for the first time, that dream reports can incorporate an experimental learning task in all phases of sleep—sleep onset, NREM and REM.

These task-related dreams could reflect neural “replay” of the learned task. But even these “direct” incorporations rarely resembled an exact replay or rehearsal of the maze, similar to what one might perform when awake. Instead, participants described their dreams as being about or taking place in a maze, with the details of the report differing from the original experience in almost every detail. Thus, it is difficult to interpret these experiences as anything resembling a “rehearsal” of learned information, and it thus is unlikely that the increased performance improvement is a result of conscious rehearsal of the route taken to reach the goal. Instead, we suggest that maze-related dreams predict subsequent task performance because they are indicators that performance-relevant components of task memory are being reactivated in the sleeping brain, albeit outside of the dream experience. It has been argued that because they fail to veridically reiterate prior experience, dreams are unlikely to support a memory function (Hartmann, 2010; Vertes, 2004). However, even the “reactivation” of memory networks described in the rodent literature is far from a faithful reiteration of experience and can “replay” never-before-experienced paths in a learned spatial environment (Gupta et al., 2010). Thus, the dream reports described here could function to link specific recent experience of the maze task with subjects’ preexisting knowledge structures about the hedge mazes, coliseums, or rooms described in the reports. Indeed, it has become increasingly clear that sleep can transform memory traces over time and brain regions, allowing for the extraction of generalizations, and integration of new information with past experience (Lewis and Durrant, 2011; Payne et al., 2009).

We also found that thinking of the task during post-training quiet rest predicted performance improvement following the subsequent night of sleep. This seemingly contradicts our prior study, in which waking thoughts of the maze were unrelated to performance following a 5hr waking delay in a control group (Wamsley et al., 2010b). This could be attributed to the fact that in our previous study Wake subjects did not sleep before being retested. Indeed, prior evidence suggests that task-related brain activity in humans during quiet rest may predict task improvement after subsequent sleep, but not wake (Gregory et al., 2014; Murphy et al., 2018).

Although our model proposes that dreams are a reflection of memory processing in the sleeping brain, alternative explanations of these data are possible. One alternative account is suggested by the fact that participants who dreamed about the maze performed especially poorly on the task at baseline. This poor baseline performance could point to a third variable driving both the incorporation of the task into dreaming and enhanced overnight improvement. For example, the brains of participants who struggle with completing the task prior to sleep may prioritize this task for consolidation during the night, in line with prior evidence that baseline performance level mediates selection of memories for sleep-dependent consolidation (Tucker and Fishbein, 2008; Wilhelm et al., 2012). Meanwhile, participants who found the task challenging may have been more likely to dream about the experience for other reasons, perhaps increased anxiety over their relatively poor initial performance. But the same explanation cannot hold true for incorporation of the task into waking cognition; although the appearance of the learning task in quiet rest reports was also associated with enhanced overnight improvement, poor baseline performance was not predictive of these incorporations.

It is not entirely clear why several other recent investigations have failed to detect an association between task-related dreams and post-sleep memory improvement (Schredl and Erlacher, 2010; Stamm et al., 2014; Wamsley et al., 2016). But in some of these studies the rate at which the learning task was incorporated into dreaming was very low, which could have prohibited a confident test of the hypothesis that these incorporations relate to subsequent memory performance (Schredl and Erlacher, 2010; Stamm et al., 2014; Wamsley et al., 2016). Indeed, in one of these studies, only one subject reported a dream that directly incorporated the learning task (Schredl and Erlacher, 2010), and in a second, only 7 task-related reports were obtained (Wamsley et al., 2016). The characteristics of learning paradigms and participant samples that might cause incorporation rates to vary so widely across studies remain unknown. At the same time, it is also likely that the true effect size of the association between dreaming and memory performance is smaller than we originally reported. Although the effect size seen here (d=1.16 for Completion Time) is still large, it is considerably smaller than the value of d=2.2 reported in (Wamsley et al., 2010b), raising the possibility that these relatively small sample-size studies were underpowered to detect this effect.

In summary, we have replicated the observation that dreaming of a recent learning experience is associated with enhanced next-morning performance. These subjective reports provide unequivocal evidence that the memory of the maze learning task was reactivated during post-training sleep. This observation, and especially the correlation of these dreams with next-morning improvement, is consistent with reactivation-based models of memory consolidation during sleep. Together with other evidence, these findings offer added support for the claim that the activation of recent memory networks during sleep is associated with subsequent performance improvement.