Memory processes during sleep: beyond the standard consolidation theory

Abstract

Two-step theories of memory formation suggest that an initial encoding stage, during which transient neural assemblies are formed in the hippocampus, is followed by a second step called consolidation, which involves re-processing of activity patterns and is associated with an increasing involvement of the neocortex. Several studies in human subjects as well as in animals suggest that memory consolidation occurs predominantly during sleep (standard consolidation model). Alternatively, it has been suggested that consolidation may occur during waking state as well and that the role of sleep is rather to restore encoding capabilities of synaptic connections (synaptic downscaling theory). Here, we review the experimental evidence favoring and challenging these two views and suggest an integrative model of memory consolidation.

The standard model of long-term memory formation

Although we spend about one-third of our life sleeping, the function of sleep is still a matter of intense debate. Obviously, sleep is an “offline” mode of the organism when it is mostly devoid of sensory information processing and goal-directed behavior. It has long been suggested that sleep contributes to memory formation by consolidating new information and by integrating it with previously stored contents [1, 2]. More specifically, both declarative and non-declarative forms of memory formation appear to benefit from sleep. Declarative (hippocampus-dependent) memory formation occurs in at least two subsequent stages [3–9]. Initial encoding of declarative memories depends on the integrity of the hippocampus [10, 11] and is probably linked to fast synaptic plasticity which leads to the formation of transient neuronal assemblies. The temporal coupling of activity between these selected neurons constitutes an early representation of experience. Encoding of new information occurs mainly during waking state and, more specifically, during exploratory behavior. The second, subsequent step of memory formation is called consolidation. During this phase, activity patterns of previously formed assemblies are reverberated and propagated into neocortical areas, probably leading to a gradually increasing contribution of the neocortex to long-term information storage. It should be noted that it is still an open question whether remote episodic memories become independent of the hippocampus, as proposed by some studies [12], or whether the hippocampus remains important for episodic aspects of memories, as suggested by the multiple trace theory [13]. This question is beyond the scope of the current review. In both cases, as a result of consolidation, memories are embedded into the network of previous knowledge and become resistant to interference [14].

In this review, we will follow the general concept of two-step memory formation. We will, however, challenge its close connection to sleep. According to the “standard model” of sleep-related memory formation which we will discuss here, consolidation occurs predominantly during sleep. For example, David Marr [15] states that during memory consolidation, “[…] ordinary sensory information must be rigorously excluded. The only time when this exclusion condition is satisfied is during certain phases of sleep.” (p. 215). Later, Gyorgy Buzsaki proposed a specific neuronal pathway for information transfer from the hippocampus to the neocortex: “In the awake brain, information about the external world reaches the hippocampus via the entorhinal cortex, whereas during sleep the direction of information flow is reversed: population bursts initiated in the hippocampus invade the neocortex.” ([4], p. 17). This model prompts three major hypotheses: (1) sleep facilitates memory consolidation on a behavioral level. As memory consolidation is operationally defined in terms of experimental psychology, the role of sleep for memory consolidation should be behaviorally measurable as increased or altered recall functions after relevant periods of sleep. (2) The second statement concerns the neural representations of experience. Because memory consolidation is related to a reactivation of experience-specific activity patterns, the standard consolidation model claims that this reactivation should occur predominantly during sleep. (3) During sleep, information should be transferred from the hippocampus to the neocortex. Thus, activity should start within the hippocampus and should be systematically followed by subsequent specific activity patterns in neocortical circuits.

Here, we will review the evidence for each of these three hypotheses. It will be shown that the first and second hypotheses are supported by some studies, while there is currently no firm evidence for the third hypothesis. Available data do, rather, support a reciprocal dialogue between the neocortex and the hippocampus during sleep-related memory consolidation, consistent with parallel processing of memory in and between both structures.

An alternative view to the standard model of sleep-related consolidation proposes that sleep serves predominantly to reduce synaptic weights and to allow for “fresh” encoding of new information after sleep [16]. This “synaptic downscaling” theory is consistent with several experimental findings: Memory consolidation is not strictly bound to sleep, but occurs constantly, also during waking state (modified hypothesis 1*); newly acquired, specific patterns of activity are not only re-played during sleep, but also during waking state (modified hypothesis 2*); and the predominant direction of propagating activity during sleep is actually from the neocortex to the hippocampus and not vice versa (modified hypothesis 3*). Importantly, however, we do not claim that the “standard model” of sleep-related memory consolidation is incorrect, but rather that it requires combination with the synaptic downscaling theory to account for the full range of available data.

Why do we need multiple long-term memory systems?

Two-step theories of memory formation suggest that an initial learning phase results in transient and labile neural representations, and is then followed by a subsequent step of memory consolidation. The function of consolidation is to stabilize engrams and to integrate freshly acquired information with previous experiences. This theory addresses the fundamental challenge that memory should both reliably store information and be plastic to allow for the modification of existing memories by new experiences. One of the first metaphorical descriptions of these requirements was given in Freud’s model of the “Wunderblock” [17]. This apparatus consists of a superficial wax layer, in which new information can be rapidly inscribed but which can also be easily overwritten, followed by an intermediate layer and a deep layer in which all events leave their enduring traces. As a result, recent experiences are immediately visible but may still be erased; their transformation into permanent representations protects them, but impedes direct access. Later, two-step memory theories have been grounded in theoretical and experimental research on specialized memory systems. David Marr [1] was one of the first to postulate an intermediate memory storage system (between short-term retention and permanent storage) on theoretical grounds. He suggested that the main function of the neocortex is to classify events based on their similarity in order to represent statistical properties of the external world, such that the connections between neocortical cells map the probability space of the environment [15]. This map should allow for the incorporation of novel experiences. However, these novel experiences need to be properly related to previous knowledge. Marr argued that assigning new experiences to their correct location within the overall probability map is computationally very complex. Integration, therefore, cannot be performed on-line, but requires an intermediate buffer which transiently stores information and relates it to previous experiences. The properties of this intermediate storage system matched what was known about the hippocampus at that time (e.g., the association network in the CA3 region) and correctly predicted some of its physiological properties which were only confirmed many years later (e.g., that the CA3 region corresponds to a random connectivity matrix). Later, this theoretical model was complemented by connectionist accounts (most prominently: [5]) which aimed at explaining the pattern of neuropsychological findings in patients with hippocampal lesions [10, 11]. In these models, rapid learning of novel associations (e.g., A–C) in the presence of existing associations (A–B) may lead to catastrophic interference, resulting in degradation of the previous connections. This effect can most easily be avoided “if new information is added gradually, interleaved with ongoing exposure to other examples from the same domain of knowledge ([5], p. 433)”. More generally, these models show that slow learning rates optimize the representation of statistical relationships by graded incorporation of new data and previous experiences. This hypothesis is highly compatible with the two-step model of memory formation where the hippocampus acquires environmental data online, while the neocortex receives information offline from the hippocampus at later stages (Fig. 1).

Fig. 1.

Two-stage theory of memory consolidation. Top during initial encoding, new information is processed (via sensory neocortex) to the hippocampus, where it is transiently stored via fast synaptic plasticity. Bottom during consolidation, the neocortex receives information from the hippocampus for permanent storage of information (figure similar to [9]). Schematic synapses depict learning-related plasticity

In the following, we will discuss the three major hypotheses derived from the standard model of memory consolidation (see above) and review the available evidence about the role of different sleep phases and of the cortico-hippocampal dialogue.

Standard hypothesis 1: memory consolidation occurs predominantly during sleep

The first hypothesis characterizing the “standard consolidation model” states that with regard to behavioral measures, memory consolidation occurs predominantly during sleep. Indeed, there is ample evidence that sleep supports both consciously accessible (declarative) and implicit memory consolidation. Moreover, some studies suggest that different stages of sleep support memory consolidation for different types of memory: while REM sleep appears to be particularly relevant for the consolidation of procedural and other non-declarative forms of memory, NREM sleep is rather beneficial for declarative memory consolidation. However, other studies question this seemingly clear distinction.

Procedural learning results in accelerated and more accurate conduction of perceptual and motor tasks. Execution of these tasks usually improves during a period of several hours following practice [18]. If subjects sleep during this period, there is an additional boost of performance in, for example, mirror-tracing [19], visual discrimination [20, 21], and execution of motor sequences (e.g., finger tapping; [22, 23]). Also other non-declarative kinds of memory have been shown to benefit from sleep (e.g., word-stem priming; [24]). It has been suggested that the consolidation of perceptual learning depends specifically on REM sleep [25]. On the other hand, some studies question the view that REM sleep is specifically related to consolidation of non-declarative memories. First, consolidation in a visual discrimination task depends not only on the second part of the night, which is dominated by REM sleep, but even more strongly on sleep in the first part of the night, which is dominated by slow wave sleep [19, 26]. Second, pharmacological suppression of REM sleep by antidepressant drugs actually enhanced the accuracy of finger tapping [27], consistent with clinical observations of preserved procedural memory in depressed patients receiving these substances over a long time period [28].

Besides procedural learning, declarative memory formation does also benefit from subsequent sleep. For example, the ability to recall vocabulary was improved by sleep after learning [29], and it was shown that sleep facilitates associative spatial learning [30]. The specific influence of different sleep states on declarative memory consolidation can be tested due to the dominance of NREM sleep during early night phases and REM sleep during later phases (reviewed in [31]). As a result, NREM sleep has been found to be particularly important for declarative memory [19, 24, 32] while procedural (implicit) memory consolidation benefits mostly from REM sleep [19, 24, 26, 33]. Furthermore, a brief day-time nap containing only NREM periods selectively enhanced memory in a paired associate task, but not in procedural learning paradigms [34]. If, in contrast, similar naps contained periods of REM sleep as well, procedural memory was enhanced [35]. Even “ultra-short” naps of only six minutes duration showed an improvement on recall of word lists [36]. On the other hand, it was found that consolidation of strictly episodic memories (located in space and time and associated with autonoetic consciousness) is more susceptible to deprivation of REM sleep than of slow wave sleep (SWS: deep stages of NREM, i.e., stages 3 and 4) [37]. Even more complicated, several studies suggest that procedural learning is improved together by SWS in the first quartile of the night and by REM sleep in the 4th quartile of the night [38].

Several mechanisms have been suggested which might account for the beneficial effect of NREM sleep on memory consolidation. During waking states and REM sleep, levels of acetylcholine are high, impairing the output from the hippocampus to the neocortex [39]. At the same time, recurrent excitatory connections within hippocampal and neocortical circuits are suppressed, favoring the processing of afferent information. During SWS, acetylcholine levels are low, thus reducing the impact of afferent synapses and facilitating internal re-play of pre-established sequences [40]. Consistent with this hypothesis, declarative memory consolidation was suppressed when healthy human subjects received cholinesterase blockers during early night [41]. An additional mechanism facilitating memory consolidation during deep stages of NREM sleep may be constituted by sleep spindles of 10–16 Hz [42] which are generated in thalamocortical circuits and promote synaptic plasticity [43]. Interestingly, sleep spindle activity is enhanced by learning [44, 45] and correlates with the formation of associations in the sensory cortex [46]. Lastly, slow oscillations <1 Hz defining responsive “up”- and silent “down”-states of the neocortex during SWS [47] may organize the bi-directional cortico-hippocampal dialogue during SWS. Artificial application of such slow alternating voltage shifts to the brain during early sleep phases can indeed facilitate declarative memory consolidation [48]. Hippocampal activity patterns involved in spatial memory formation and the cortico-hippocampal dialogue during memory consolidation will be discussed in the next section on the Standard hypothesis 2.

Taken together, these data provide ample evidence for the facilitating role of sleep in memory consolidation. However, memory processing may be improved not only by subsequent sleep, but also by sleeping prior to learning. Yoo et al. [49] observed that one night of sleep deprivation significantly impaired the encoding of new episodic memories and was associated with reduced hippocampal activation during encoding. Similarly, Van der Werf et al. [50] found that even a reduction of sleep depth deteriorated subsequent hippocampus-dependent memory formation. We recently tested this idea using a recognition memory paradigm in epilepsy patients with intracranial EEG electrodes in bilateral hippocampi and rhinal cortices [51]. Subjects were presented pictures of buildings and landscapes on each of two subsequent days; on each day, a recognition memory test followed. One day contained a nap of 1 h duration between the first and second learning session, while the other day contained a period of the same duration without sleep. A schematic depiction of this paradigm and the main results is provided in Fig. 2. While we did not observe an effect of sleep on retrieval of information acquired before sleep, retrieval of the items learned after sleep was significantly improved. It should be noted that the same paradigm has been conducted using fMRI [52]. In this study, naps affected memory for items encoded prior to sleep, consistent with the standard consolidation theory. This apparent discrepancy might, however, be explained by different recording modalities and experimental parameters (e.g., testing after different time intervals in the study by Takashima et al. [52] as compared to our study [51]). Both studies report increased reaction times for remote as compared to recent items (Fig. 2d). Such an increase in reaction times following sleep might seem to indicate less consolidation, but this is not necessarily the case: at least for declarative memory, consolidation implies that memories are embedded into larger networks of previous experiences, which might slow down rapid access to these memories [53]. Using intracranial EEG, we found that event-related hippocampal potentials on the control day were more negative during retrieval of recent than remote items (Fig. 2e, f). Such an increased negative potential likely corresponds to an enhanced activation of this region (e.g., [54]), which in turn indicates that these memories have undergone consolidation to a lesser degree at the time of retrieval. On the day with nap, in contrast, there was no difference in event-related potentials during retrieval of recent and remote items, suggesting that the nap facilitates at least early stages of memory consolidation of the recently learned items. These findings indicate that sleep can prepare the brain for better memory consolidation which can, however, itself occur in wake states (modified hypothesis 1*, see above). They are also consistent with the suggestion that one function of sleep is to reduce synaptic weights and thereby to restore the brain’s encoding capacities [16], but move beyond it by suggesting that not only memory encoding, but also consolidation is facilitated by prior sleep. Before we describe this theory in greater detail, we will turn to the evidence for and against the two other hypotheses derived from the standard consolidation theory.

Fig. 2.

Experimental design and results from a study of memory consolidation during a brief afternoon nap. a Experimental paradigm. b According to the Standard model of memory consolidation, consolidation occurs predominantly during sleep. c According to the alternative view, consolidation may also occur during waking state, but is facilitated after sleep. d–g Reaction times and intracranial event-related potentials (ERPs) within the hippocampus are consistent with the alternative model (modified from [51])

Standard hypothesis 2: memory representations are predominantly reactivated during sleep

The second hypothesis claims that the re-activation of neuronal assemblies (the proposed neuronal substrates of memory representations) occurs mainly during sleep. Before reviewing the respective literature, we will briefly describe how such sequences arise and how they may be related to memory. In rodents, individual cells in different subregions of the hippocampus increase their firing rate if the animal is located at a specific spatial location and have thus been termed “place cells” [55, 56]. During spatial exploration, action potentials of these neurons are entrained by the underlying theta network rhythm [57–60]. When the animal crosses the respective place field of a given place cell, the phase-relationship to the theta oscillations systematically proceeds, thereby generating a sequential activation of place cells with overlapping place fields. In principal, these sequences allow for the reconstruction of the animal’s trajectory through the explored territory. Thus, by definition, they form a temporal representation of the navigated space. Later, Wilson and McNaughton [61] discovered that sequences of action potentials in place cells with overlapping place fields are re-played, in a temporally condensed manner, during sleep. They occur on top of hippocampal sharp wave-ripple complexes [62]. Sharp waves are field potential waves of 30–200 ms duration which are generated in the hippocampal subfield CA3 and propagate through the hippocampal output loop into the entorhinal cortex [63]. In CA1, they are superimposed by very fast network oscillations, reaching more than 200 Hz in rodents, termed ripples [62]. Together, theta-phase precession during spatial exploration forms temporal sequences of activity and sharp wave-ripple complexes provide a scaffold for their re-play during sleep. Indeed, sequence replay was observed both during SWS periods [61, 64–66] and during REM sleep [67]. Even sequences of three spikes fired by three different place cells during wheel-running are repeated in subsequent slow wave sleep [68], and information about the order in which the cells were activated is maintained [69]. These findings support the idea of temporal coding in transiently formed neuronal assemblies as a mechanism of experience-dependent learning [70]. The occurrence of sharp wave-ripple complexes during sleep or awake immobility [62] and their propagation into the adjacent entorhinal cortex [63] does also lend strong support to the two-stage model of spatial memory formation.

Replay of sequences is not restricted to the hippocampus, but has been observed in other regions as well: Behavior-related correlations between hippocampal and neocortical activity again occur during subsequent SWS [71, 72]. Correlations of cell-pair activity in the medial prefrontal cortex which were established during waking state were also observed in the subsequent night [73]. Ribeiro et al. [74] even described replay-related activity patterns during SWS in hippocampus, neocortex, putamen and thalamus, although this study was later criticized for methodological reasons [75].

Taken together, these studies provide good evidence for a re-activation of both hippocampal and neocortical activity patterns during sleep. However, in contrast to the canonical two-stage model of memory formation, ripples and ripple-coordinated neuronal discharges are not restricted to sleep (modified hypothesis 2*). During exploratory behavior, theta oscillations are not continuous, leaving room for the occurrence of sharp wave-ripple complexes [76]. During such ripples, place cells with similar or overlapping place fields increase their correlated firing during sharp waves and maintain this coupling during subsequent sleep-related sharp wave-ripple complexes. Ripples occurring during exploratory behavior may indeed be involved in the generation of strong and persistent temporal coupling between hippocampal pyramidal neurons. These cells show an increased amount and temporal precision of coupling during ripples which occur in unfamiliar environments as compared to familiar environments [77]. Conversely, repeated activation of assemblies during behavioral repetitions in rats potentiates the occurrence of sharp waves and re-activated assemblies [78]. It is therefore quite feasible that fast ripple oscillations play an active role in the induction of plasticity processes which stabilize place-encoding neuronal assemblies. Indeed, sharp waves are prone to induce long-term potentiation in hippocampal CA1 pyramidal neurons [79] and sharp wave-ripple complexes in brain slices in vitro can be generated by plasticity-inducing stimulation paradigms [80].

Thus, the formation of assemblies by sharp wave-associated co-activation does already take place during wakefulness. Sequential activation of place cells on top of high-frequency oscillations has also been found during brief resting periods in rats running back and forth on a linear track. These sequences resembled the activation of place cells along the track but, surprisingly, occurred in reverse order [81]. Again, the underlying field activity consisted of sharp wave-ripple events. During subsequent sleep periods, re-activation was forward with respect to previous experience on the track. Later, this finding has been confirmed but replay occurred more frequently in the forward direction than in reverse order [82]. The inverted order of place cell discharges was mostly present immediately after a run through the track whereas forward re-play was found before running. It is feasible that these different modes of re-play are related to the influence of recent experiences and anticipations of future activities, respectively [83, 84]. While these findings demonstrate that replay may also occur during waking state, they do not exclude the possibility that replay which is important for consolidation (i.e., which supports transfer of a memory trace from the hippocampus to the neocortex; see next section) predominantly occurs during sleep.

Although hippocampal ripple activity during SWS increases after learning [85], it should be noted that we are still lacking a proof for a causal link between sequence replay during sleep and memory consolidation. One way to tackle this question is to correlate the time and quantity of sequence replay with behavioral recall after sleep. The subsequent memory tasks can be most easily performed in human subjects; however, identification of stimulus-specific neural activity in humans is difficult. Single-unit recordings from the hippocampus of epilepsy patients revealed that individual cells represent specific items [86], but the apparent lack of a topographical organization of these representations suggests that field potentials (which can be recorded with standard macroelectrodes in these patients) are not stimulus-specific [87, 88]; for review, see [89].

However, several studies on humans describe a reactivation of memory-related brain regions during sleep following learning periods (hypothesis 2). Using PET, Maquet et al. [90] found that brain regions which were activated during a serial reaction time task show enhanced metabolism during REM sleep in the consecutive night. After declarative learning, the hippocampus is increasingly activated during periods of SWS, as shown by fMRI [91, 92]. Again, however, there is also evidence that replay is not restricted to sleep, but can occur during waking states as well (modified hypothesis 2*). In one study, Peigneux et al. [93] investigated BOLD activity in fMRI immediately after a hippocampus-dependent declarative learning task or a hippocampus-independent serial reaction time task and observed task-specific offline reactivation of the respective memory systems during waking state. In a second study, Gelbard-Sagiv [94] recorded single neurons within the hippocampus of epilepsy patients. These cells responded selectively to a specific famous person during a learning session and were again active prior to free recall of this person. However, in both studies it is unknown whether the observed reactivation patterns actually correspond to memory consolidation; especially in the study by Gelbard-Sagiv where replay did not occur spontaneously but was triggered by retrieval effort.

Do ripples occur in human subjects, and are they closely linked to sleep? In general, data from the human hippocampus are difficult to obtain because of its deep location and specific field properties. Therefore, it is not possible to record hippocampal ripples via scalp EEG. However, epilepsy patients with intracranial electrodes implanted during presurgical investigations allow studying human hippocampal activity. Ripples with a maximum power between 80 and 140 Hz were detected in the human hippocampus and rhinal cortex using microelectrodes with a diameter of 40 μm [95–97]. In addition, faster oscillations with a power maximum between 250 and 500 Hz (termed “fast ripples”) were observed in close proximity to the epileptic focus and thus appear to be related to pathological processes [95–98]. Ripples with a slightly lower frequency maximum were also observed with macroelectrodes, which are routinely used for clinical purposes [99, 100]. Furthermore, Clemens et al. [101] recorded ripples using foramen ovale electrodes with a tip located in the subdural space close to the parahippocampal gyrus.

Interestingly, in several of these studies ripples were observed not only during sleep, but occurred during waking state as well: while Staba et al. [102] found that the rate of ripples during NREM sleep was twice as high as during waking state, Clemens et al. [101] observed that more ripples occurred during stage 1 and 2 of NREM sleep as compared to waking state, but that the incidence of ripples during SWS and waking state was comparable. Similarly, Bagshaw et al. [103] reported hippocampal ripples and fast ripples during all states of vigilance, including REM and wakefulness. Again, the occurrence of the fast events was most pronounced during stage 1–2 of NREM sleep. We found that the rate of ripples during waking state was even significantly higher as compared to all sleep stages [100]. It should be noted that the term “waking state” usually corresponds to normal awake behavior, whereas in our study, it referred to periods before or between sleep phases when subjects lay in a quiet dark room with their eyes closed. Thus, this state is more closely related to a resting state of the brain than to awake exploratory behavior. Possibly, replay during this awake resting state plays a similar role for memory consolidation (i.e., supports transfer of information from the hippocampus to the neocortex) as replay during sleep, although there is currently no direct evidence for hippocampal-neocortical cross-talk during this state. Together, these data strongly support the idea that not only reactivation of memory systems, but also ripple activity is not restricted to sleep, but occurs during waking states as well (modified hypothesis 2*).

Standard hypothesis 3: during sleep, information is transferred from the hippocampus to the neocortex

The third hypothesis proposes that during sleep, hippocampal markers of reactivation occur time-locked to, but briefly earlier than, neocortical activity patterns related to information encoding. This temporal order would be indicative of information transfer from hippocampus to the neocortex. It should be noted that thus far, the neural mechanisms supporting permanent storage of information are still unclear (see next section). However, several studies observed that neocortical sleep spindles, which occur with the highest incidence during stage 2 sleep, and slow waves depend on previous learning episodes. Again, this issue was investigated both in animals and humans. In humans, performance of a declarative memory task enhanced the density of sleep spindles in the subsequent night [104]. Parahippocampal ripples recorded with foramen ovale electrodes were locked to sleep spindles in surface EEG recordings [101]. These studies are, in principle, consistent with information transfer from the hippocampus to the neocortex, although the direction of coupling was not investigated.

The coupling of hippocampal ripples to neocortical sleep spindles and slow waves has also been studied in rats. Indeed, hippocampal ripples are temporally linked and mostly precede neocortical sleep spindles [105]. Furthermore, hippocampal ripples occurred time-locked to neocortical sleep spindles [106, 107]. In vivo recordings also revealed that hippocampal sharp waves are locked to the depolarizing phase of neocortical delta waves/slow rhythm, which is associated with enhanced cortical activity [106–109]. Likewise, sleep spindles and ripples are concentrated on the depolarizing phase of neocortical slow activity [110], suggesting that these coordinated patterns of activity take place during a highly responsive and plastic state of the neocortex [42, 111].

Again, while these studies show that hippocampal ripples, neocortical sleep spindles and up-states are locked in time, they do not reveal the direction of the hippocampo-cortical or cortico-hippocampal information transfer. Indeed, during SWS, hippocampal activity is influenced by neocortical transitions between up- and down-states, while an impact of the hippocampus on neocortical activity has not been observed [109]. Up to now, there is no clear evidence that the hippocampus specifically influences the neocortex during sleep (see the comment by Tononi et al. [112]). While Siapas and Wilson [105] had established a correlation between hippocampal ripples and neocortical sharp waves, a closer investigation of action potential timing revealed that neocortical spikes related to spindle and delta activity precede ripple-related discharges in the hippocampus by about 50 ms [106]. These findings are consistent with information transfer from the neocortex to the hippocampus, but not in the reverse direction. It may be argued that hippocampal activity always follows activity in the neocortex, but this is not true either: during waking state, prefrontal activity is locked to hippocampal theta oscillations [113], suggesting that the hippocampus actually exerts an influence on neocortical activity.

We recently investigated this question in human intracranial EEG recordings (unpublished observations). A directional coupling analysis [114] was used to study interactions of rhinal cortex, hippocampus and neocortex during waking state and during different sleep stages. While in general the predominant coupling direction was from the anterior hippocampus to the rhinal cortex and neocortex, the neocortex exerted an increasing influence on the hippocampus during sleep. Taken together, several studies indicate that hippocampus and neocortex interact closely and specifically during sleep. The predominant direction of influence, however, appears to be from the neocortex to the hippocampus and not vice versa (modified hypothesis 3*).

Synaptic potentiation or depression during sleep?

If, according to the standard consolidation theory, consolidation occurs predominantly during sleep, this state should also facilitate the cellular correlates of learning. One of the most important candidates for such a correlate is long-term potentiation (LTP), although this relation still remains disputed (see below). Such an “extended” standard consolidation theory would also predict a link between sleep and LTP. On the other hand, one of the most specific predictions of the synaptic downscaling theory is that sleep corresponds to a net depression of synaptic weights. In this section, we will review studies attempting to relate synaptic plasticity and sleep. Several studies linked sleep spindles with LTP. First, using computer simulations, Sejnowski and Destexhe [115] suggest that sleep spindles play a crucial role for the induction of long-term plasticity in the neocortex by triggering Ca²⁺ entry through dendritic depolarisations. Second, the local spindle density correlates with efficacy of TMS-induced spike-timing dependent plasticity (as measured by motor-evoked potentials after stimulation of both the median nerve and the hand region in the primary motor cortex; [46]). Finally, hippocampal ripples were related to LTP as well [79, 80]. However, there is also evidence relating sleep-related activity and synaptic depression. First, Colgin et al. [116] found that the induction of LTP was impaired in slices which showed spontaneous ripples. Second, Vyazovskiy et al. [117] observed a net synaptic potentiation during waking state and a depression during sleep. Third, sleep deprivation was found to impair LTP induction, possibly due to saturation [118, 119]. These observations support the synaptic downscaling theory, which suggests that the function of sleep for memory formation is to restore synaptic weights to allow for the efficient encoding of new information after sleep. This reduction of synaptic weights should improve the signal-to-noise ratio of cortical information processing because weakly potentiated synapses corresponding to interfering noise are completely silenced and only strongly potentiated synapses corresponding to the signal survive. More specifically, the synaptic downscaling theory (or “synaptic homeostasis hypothesis”; [16, 120]) is based on two observations: First, that the intensity of slow waves depends on the amount of accumulated prior waking time [121]; and second, that acquisition of new information during waking state is linked to sub-cellular modifications which eventually lead to an increase of synaptic weights, i.e., to synaptic potentiation [122].

Despite its role for learning in simple organisms such as aplysia and its existence in the mammalian neocortex (e.g., [123]), the exact functional role of LTP for learning and memory in mammals has still not been identified. One important form of LTP, which may serve as an implementation of the Hebbian learning rule of coordinated pre- and postsynaptic neuronal activity [124], depends on activation of N-methyl-d-aspartate (NMDA) receptors and a subsequent cascade of molecular and genetic changes [125, 126]. For example, it was shown that hippocampus-dependent spatial learning in mice is impaired after spatially selective deletion of the NMDA receptor in the CA1 region of the hippocampus ([127]; mice completely lacking the NMDA receptor die postnatally), or in mice lacking the alpha-calcium-calmodulin kinase II (alpha-CaM-Kinase II; [128]). Similar deficits were induced by pharmacological blockade of NMDA receptors via intra-hippocampal infusion of APV [129]. These and other putative molecular correlates of learning (such as BDNF, CREB, and Arc) are up-regulated during waking state, e.g., due to an increased activity of noradrenergic neurons [130, 131]. The synaptic downscaling theory predicts that the net synaptic potentiation during waking state is balanced by slow wave activity during sleep. Indeed, animals with lesions of the noradrenergic system and a subsequent reduction of LTP-related gene expression show a decrease in slow wave activity [130, 131]. In humans, putative induction of regionally-specific neural plasticity is associated with a subsequent enhancement of slow wave activity in the same brain area [132], while arm immobilization causes a reduction of local slow wave activity within the sensorimotor cortex [133]. Several results suggest that slow wave activity actually promotes a depression of synaptic weights. First, a frequency of ~1 Hz is ideal to reverse the effect of NMDA receptor dependent LTP by removing AMPA receptors from the postsynaptic membrane (e.g., [134]). Second, the reduced concentration of neuromodulators such as acetylcholine (e.g., [135, 136]) and noradrenaline [137] is favorable for a reduction of synaptic weights. In addition, substances such as insulin and calcineurin which promote synaptic depression are up-regulated during sleep [138, 139].

It is less clear whether neocortical information storage depends on similar neural mechanisms. Some studies indicate that retrieval of remote memories enhances the expression of immediate early genes [140, 141] and depends on the alpha-CaM-Kinase II [142], which is closely linked to induction of LTP. Furthermore, it was shown that NMDA receptor expression is required for retrieval of old memories [143], although the neocortical localization of these receptors was not directly demonstrated in that study. Based on these studies, it has been suggested that long-term storage of information in the neocortex actually depends on LTP [9]. Importantly, however, these studies did not investigate the circadian effects on neocortical plasticity. Furthermore, several alternative plasticity mechanisms may account for long-term memory encoding as well. For example, long-term depression (LTD) was associated with exploration of novel environments in rats [144] and may also be important in the neocortex to sharpen activity in neural assemblies during learning [88, 145]. In addition, non-synaptic plasticity mechanisms (for reviews, see [146, 147]) might also play a role during neocortical memory formation, in particular because spindle-related activity was shown to control dendritic excitability [148].

Conclusion

Above, we have evaluated three major hypotheses underlying the standard model of memory consolidation (see Table 1). Indeed, a large amount of data indicate that memory replay and consolidation take place during sleep, in particular during NREM sleep. However, several studies have also provided evidence that reactivation of memory representations and consolidation occur during waking states, contributing to the formation of neuronal assemblies. In contrast to common concepts of memory formation, sleep is associated rather with neocortical LTD than with LTP. Furthermore, studies investigating the interaction between hippocampus and neocortex do not support the idea of a hippocampal–neocortical information transfer during sleep, but rather suggest a co-activation or even a neocortical driving influence on the hippocampus.

Table 1.

Overview of the experimental evidence for the “Standard consolidation theory”

	Standard consolidation theory	Alternative theory
Function of sleep	Consolidation of previous experiences	Synaptic downscaling for consecutive learning
Memory consolidation	NREM: Declarative memory REM: Procedural memory	Continuous (during resting states immediately after encoding)
Memory replay	During sleep	During sleep and waking state
Neocortical LTP during sleep	Increase (if encoding is related to LTP)	Decrease
Hippocampal–neocortical interactions during sleep (memory transfer)	Hippocampus → neocortex	No predictions

Predictions supported by experimental evidence are given in italics. Predictions with opposing experimental evidence are given in bold

To conclude, we think that an integrative view of the role of sleep for memory consolidation, which combines (slightly modified) elements from the standard theory of memory consolidation with ideas from the synaptic downscaling theory, is most consistent with the experimental evidence at hand: Such an integrative view comprises the idea that replay of activity sequences and memory consolidation occur during resting states characterized by decreased sensory and cognitive processing. Although NREM sleep is particularly suited for memory replay and consolidation, those processes, in principle, may start immediately after memory encoding. Furthermore, this view implies that sleep causes a downscaling of neocortical synapses. This downscaling promotes memory consolidation by refinement and sharpening of previously acquired memories and an increase of the signal-to-noise ratio. Finally, synaptic downscaling improves learning after sleep by preventing saturation of synaptic weights.