An opportunistic theory of cellular and systems consolidation

Introduction

One of the central questions in cognitive neuroscience is: how does the human brain learn new information while not overwriting previously stored memories? Memory models have addressed this so-called stability plasticity dilemma [1] by proposing a post-encoding, offline period that allows the brain to consolidate recent experience into long-term storage [2–4]. Recent models propose that this offline period takes place during sleep [5]. In general agreement with this idea, a large number of behavioral studies have demonstrated that sleep facilitates both non-declarative memories (i.e., non-conscious memories such as sensorimotor skills, perceptual skills, habits, priming, etc.) [6–11], which are not dependent on the hippocampus, and declarative memories (i.e., conscious memories of events, places, and general knowledge), which are dependent on the hippocampus. Although both kinds of memory benefit from sleep, they benefit in different ways. Non-declarative memories are often enhanced after a period of rapid eye movement (REM) sleep (Glossary) such that performance increases from the end of training [6, 8, 12–15]. In contrast, the sleep-related facilitation of declarative memories usually consists of decreased forgetting after a post-encoding period of non-REM (nREM) sleep composed of Stage Two and Slow Wave Sleep (SWS)[10, 16–19].

These opposing performance outcomes (i.e. enhanced memory performance versus decreased forgetting; Figure 1), raise the possibility that the consolidation of declarative and non-declarative memories rely on distinct mechanisms. With regard to declarative memory, which is the main focus of our Opinion article, we ask the question: What are the essential conditions for consolidation to occur? It has previously been proposed that SWS is critical for declarative memory consolidation [20, 21]. The viewpoint that we advance here suggests that the critical condition is a period of reduced interference, during which consolidation renders the memory trace less vulnerable to the negative effects of new encoding. In addition, we argue that this process occurs as a result of both cellular consolidation and systems consolidation.

SWS facilitates declarative memory

A long history of research has consistently shown that a period of SWS yields less forgetting than a comparable period of REM sleep or waking activity [9, 10, 17, 22]. In humans, the first half of the night is rich in SWS, while the second half of the night is rich in REM sleep. Previous studies have shown that 4 hours of mostly SWS resulted in less forgetting of a declarative verbal memory task performed earlier in the day than a comparable period of REM sleep or waking, which did not differ from each other in terms of performance [9, 22]. Similar results were found in a study that controlled for time-of-day and circadian confounds[9]. Subsequent studies replicated the effect [10, 17]. Similarly, compared with wake, daytime sleep (naps) (consisting of only nREM) showed less forgetting on a declarative paired word associates task, but not on a non-declarative procedural memory task. Furthermore, better performance in this task was positively correlated with minutes of SWS during the nap [23].

Why is SWS good for declarative memory? A number of different hypotheses to explain this have been put forward over the years. In the following section, we will briefly outline two hypotheses, namely the Unique-to-sleep consolidation hypothesis and the Passive Interference Reduction hypothesis, that have been previously put forth in the sleep field, as well as outlining the rationale behind our Opportunistic consolidation hypothesis.

Phases of declarative memory consolidation

The encoding of a memory trace begins with activation of primary and associative cortical areas. The hippocampus integrates information from these distributed cortical structures and rapidly fuses these features into a coherent memory trace. At least two phases of consolidation have been identified, which together trace the process of consolidation from the hippocampus to cortex: cellular and systems consolidation. A review of the phases of consolidation is necessary to understand the effect of SWS and other brain states on memory for previously learned information.

The role of SWS in cellular and systems consolidation

SWS has been shown to reduce experience-dependent hippocampal plasticity, thereby inhibiting the formation of new memories. For example, in sleeping rats, LTP can be readily induced during REM sleep, but is much less readily induced during SWS [59]. Similar to NMDAR antagonists, SWS inhibits the induction of LTP, but does not disrupt the maintenance or persistence of previously induced LTP [60]. In contrast, regarding synaptic plasticity in the hippocampus, REM sleep is similar to the awake state, as LTP can be readily induced [60]. Similarly, associative learning is impaired during SWS, and not during REM. For example, in a classical conditioning paradigm using brain stimulation as the conditioned stimulus (CS) and unconditioned stimulus (US), with multiunit firing of hippocampal neurons as a measure of associative learning, paired CS-US brain stimulation resulted in increased hippocampal firing in response to the CS during REM and wake, but not during SWS[61]. Thus, in rodents, it appears that SWS is a period of reduced hippocampal plasticity. As a consequence, few memories should be formed during SWS, compared with REM or wake, even though considerable mental activity occurs during SWS[62].

If memories are not encoded during SWS, then post-training SWS should be conducive to the cellular consolidation of both LTP and memories formed during training, just as post-training NMDAR antagonists are often conducive to both. LTP induced during wakefulness is later fully expressed during SWS [60], which is consistent with our proposed view. However, it is not known if SWS is more protective of previously induced LTP than a corresponding period of wakefulness or if a temporal gradient of interference reduction (with SWS being more protective of LTP soon after induction) is observed, as our hypothesis predicts. With regard to its effect on declarative memory, SWS has in fact been shown to have such effects. Sleep studies with humans have reported a temporal gradient of interference that is similar to the gradient obtained in studies of the effect of NMDAR antagonists in rats discussed earlier. Using a 24-hr retention interval between initial learning and a subsequent memory test, these studies showed that sleep soon after learning is more beneficial than sleep that is delayed[63], even after controlling for circadian confounds [64, 65]. The temporal gradient associated with sleep is consistent with the notion that reduced hippocampal plasticity protects recently formed and still-fragile memories, giving them a chance to become hardened against the forces of retroactive interference as cellular consolidation unfolds.

In addition to benefiting cellular consolidation, the absence of encoding during SWS may also be conducive to neural replay and systems consolidation of memories formed prior to sleep. Indeed, in rodents, the coordinated replay of place cell activity has often been reported during SWS [52, 54, 55]. Analogously, in a positron emission tomography (PET) study with humans, hippocampal areas that were activated during a route learning task were reactivated during subsequent SWS [66]. Interventions designed to enhance reactivation during SWS have also been shown to affect consolidation. In a functional magnetic resonance imaging (fMRI) study with humans, cuing recently formed odor-associated memories by odor re-exposure during SWS (but not during REM sleep) prompted hippocampal activation during sleep resulting in less forgetting after sleep, compared to a control group [67]. Overnight sleep increased the integration of newly learned words into the mental lexicon (i.e. lexical competition effect), a measure of systems consolidation, and the degree of lexical competition correlated with sleep spindles during NREM sleep [68]. Similarly, transcranial direct current stimulation of delta activity (<3Hz) during SWS significantly decreased forgetting in declarative memory but not procedural memory performance[69]. Thus, these results do not show that systems consolidation improves the fidelity of the memory trace (i.e. increased retention from baseline was not observed). Instead, SWS appears to protect new memories by consolidating them in a low-interference storage site (such as the cortex).

Computational models of the role of the hippocampus in declarative memory posit that the degree to which new encoding interferes with recently learning material is related to the amount of overlap in the synaptic representations of successive experiences, with greater overlap leading to greater interference [70–73]. It is therefore possible that the mechanism by which SWS improves retention during reactivation involves facilitating the consolidation of multiple memories in a more distributed, non-overlapping fashion in the cortex. If so, then the reduced interference associated with SWS may render memory traces more resistant to interference in two (related) ways: (1) by allowing the molecular processes associated with cellular consolidation to proceed undisrupted, and (2) by facilitating the activation (and reactivation) of the distributed network between hippocampus and various cortical areas through the processes associated with systems consolidation (Figure 3).

Quiet Wake: the resting mind at replay

Much of the earlier work on neural replay documented its occurrence during SWS, which was consistent with the idea that this form of consolidation may have been sleep-specific. However, recent evidence suggests that replay of recent experiences[54, 74, 75], and future experiences [76], as well as remote replay (i.e., replay of sequential place cell firing that occurred earlier in a different environment) can occur during the awake resting state as well[77–79]. These findings suggest that the hippocampus may take advantage of any down time (including, but not limited to, SWS) to consolidate memory. That is to say, the processes that underlie systems consolidation may unfold whenever the hippocampus is not encoding new memories (e.g., [53]).

Indeed, prior studies have suggested that hippocampal plasticity is reduced during quiet wake just as it is during SWS [80, 81]. Thus, we argue that reduced experience-dependent hippocampal plasticity is what allows cellular consolidation mechanisms to evolve without interference and simultaneously sets the stage for systems consolidation mechanisms to become operative. Several recent human studies have indicated that quiet wake, a non-sleep resting state with reduced encoding and interference, can induce a similar behavioral enhancement as sleep. In a study using a hippocampus-dependent visual search task [82, 83], similar learning profiles were reported for the nap and quiet wake group, but the active wake group showed less learning [8]. The authors hypothesized that quiet wake mimicked the dampened encoding found in SWS, which facilitated consolidation. Importantly, most sleep studies compare sleep with an active wake period as the control, which does not control for waking experiences that could potentially cause interference with recently encoding memories. The few studies that have compared quiet wake with sleep have reported similar memory gains in both groups [84, 85]. In a similar vein, hippocampal-cortical connectivity in humans, as measured by blood-oxygen-level-dependent (BOLD) fMRI, during quiet wake immediately following a learning task was enhanced compared to a pre-task resting baseline[86]. Further, individual differences in the magnitude of post-task functional connectivity was predictive of later associative memory performance [85]. Thus, data from both rodent and human studies are consistent with our hypothesis that hippocampal-based memory consolidation utilizes optimal brain states to process prior learning, and is not specific to sleep per se.

Pharmacologically induced SWS-like neural states

According to the Opportunistic Consolidation Hypothesis, neither SWS nor quiet rest uniquely trigger consolidation processes. Instead, they facilitate consolidation because they are both associated with reduced hippocampal plasticity (and, therefore, reduced encoding of new memories). Amnesia-inducing drugs, such as alcohol, benzodiazepines and acetylcholine antagonists, modulate memory performance similarly to sleep, perhaps also due to reduced hippocampal plasticity [87]. That is, these drugs may selectively block the encoding of new memories, which may initiate cellular and systems consolidation of memories that were encoded in the recent past.

With respect to ethanol, multiple studies in rodents and humans have demonstrated that it can impair the acquisition of new information when administered before learning [88–91]. However, when taken after learning, alcohol facilitates memory performance more than placebo, in the sense that less forgetting is observed [92, 93]. Similarly, benzodiazepines (GABA_A receptor allosteric modulators), have been shown to protect memory for information learned prior to drug administration, and hence, only induce anterograde amnesia [94–96]. A combination of scopolamine [a muscarinic acetylcholine receptor (AchR) antagonist] and mecamylamine (a nicotinic AchR antagonist) produced retrograde facilitation of word pairs compared with placebo, as well as decreasing acquisition of new words [97]. Consistent with the interference literature, the enhancement of memory consists of less forgetting in the drug group compared to placebo (not an absolute increase in performance relative to prior performance).

It seems unlikely that amnesia-inducing drugs yield less forgetting of previously encoded memories by directly enhancing the consolidation process. This is because a direct enhancement of consolidation would be expected to also yield anterograde facilitation (i.e., the enhancement of memories learned after the drug is taken), not anterograde amnesia. Instead of these drugs directly augmenting consolidation processes, we hypothesize that the mechanism by which amnesia-inducing drugs retroactively facilitate memory involves reduced experience-dependent hippocampal plasticity. The reduced encoding of new memories indirectly facilitates cellular and systems consolidation processes by allowing them to opportunistically proceed for memories that were encoded in the recent past. In agreement with these considerations, ethanol produces a dose-dependent suppression on the magnitude of LTP following high frequency stimulation in acute hippocampal rodent slices [98], but it specifically blocks the induction of LTP, and not the expression or maintenance of LTP [89]. Thus, it seems reasonable to propose that ethanol may facilitate cellular consolidation. Benzodiazepines also block the induction of hippocampal LTP [99], as do AchR antagonists [100]. Thus, using LTP as a model, alcohol, benzodiazepines and AchR antagonists would be expected to (and do) induce anterograde amnesia, which would (according to our theory) protect older memories from retroactive interference, thereby giving them a chance to undergo cellular consolidation (Figure 2).

Anterograde amnesia may also disinhibit systems consolidation processes (such as neural replay) by quieting hippocampal encoding activity and allowing for hippocampal-cortical dialog associated with previously encoded memories [53]. The results of a recent study [101], which trained rats to learn three new goal locations in an open field each day, is in agreement with this view. Goal-related place-cell replay activity during rest (sleep was not measured) generally corresponded to the newly learned locations. However, the use of an NMDAR antagonist resulted in rapid forgetting of new locations as well as place-cell replay activity that corresponded to the previous day's goal locations [101]. Consistent with these data, studies of amnesics show that delaying a post-encoding period of interference improves recall of words, compared with no delay [102].

Although parallels can be drawn between SWS, quiet wake and amnesia-inducing drugs with regard to their apparent effect on cellular consolidation, it is not yet known whether amnesia-inducing drugs also liberate processes associated with systems consolidation. For example, it is not known whether benzodiazepines and alcohol facilitate communication between the hippocampus and cortex via hippocampal sharp waves/ripples and cortical spindles. We speculate that, like SWS, amnesia-inducing drugs that selectively block the encoding of new memories may, indeed, trigger the switch between encoding and consolidation. That is, the Opportunistic Consolidation hypothesis holds that a temporary period of anterograde amnesia may be the common denominator that ties together a variety of conditions that are favorable to consolidation. Future studies are needed to test this hypothesis.

Acetylcholine: a modulator of declarative memory consolidation

What is the underlying mechanism that may be supporting the switch between encoding and consolidation? In a 2-stage model [53, 103], ACh signaling functions as a switch between inflow (encoding or write-in) and outflow (consolidating or write-out) modes of the hippocampus. In the encoding state, high cholinergic activation, such as during active awake and REM sleep [104], when theta rhythm is present [105, 106], sets the appropriate hippocampal dynamics for inflow of information by suppressing communication from the hippocampus to the cortex [107]. At low cholinergic activation, such as during quiet wake and SWS when hippocampal sharp waves are present [53], there is a release from cholinergic suppression. This permits outflow of information from the hippocampus to the cortex [103]. Thus, according to this view, the consolidating state is not specific to sleep, but it does occur during sleep. Critically, the encoding and consolidating states are also associated with characteristic rhythmic activity, and a basic assumption of this account is that communication between the hippocampus and cortex is mediated by coordinated oscillatory rhythms across different structures of the brain [108]. A similar suggestion has been made in computational models of consolidation [5].

Lower levels of ACh prevail during quiet wake and SWS [109], and this is thought to shift the hippocampus into the consolidating state [97]. In this state, activity along input pathways (ordinarily facilitated by cholinergic-driven, theta rhythm) is suppressed and hippocampal plasticity is low (i.e. hippocampal LTP is not readily induced). As such, and as indicated earlier, recently induced LTP would be protected from interference and would be given a chance to stabilize as the process of cellular consolidation unfolds. In addition, under these conditions, the cortex is characterized by low frequency spindle (i.e. 7–14 Hz) and delta (i.e. ≤ 4 Hz) oscillations [110], whereas the hippocampus is associated with a more broad-spectrum pattern punctuated by brief, high frequency sharp waves (i.e. 30+ Hz) and very high frequency ripples (∼ 200 Hz) [53]. These sharp wave oscillations occur within the hippocampal-entorhinal output network. Synchronized neural discharges tend to occur along this pathway during sharp-wave/ripple events [111, 112]. Thus, once again, rhythmic activity seems to coordinate communication between adjacent brain structures.

Such communication has been found to occur between more distant brain structures as well. For example, ripples observed during hippocampal sharp waves have been found to be correlated with the occurrence of spindles in the PFC [55]. Moreover, the neural replay discussed earlier preferentially takes place during the high-frequency bursts of spindle waves [113]. All of this suggests that rhythmically-based feedback activity from the hippocampus may serve to “train” the cortex and thus facilitate the process of systems consolidation. Importantly, this brain state is not specific to SWS but also generalizes to quiet wake as well. If the cessation of encoding triggers this ACh-modulated switch between encoding and consolidation, then we might predict that this same switch would be triggered by other states of reduced encoding, such as pharmacologically-induced states of anterograde amnesia. Interestingly, the drugs that would be expected to have this effect are those that selectively inhibit the encoding of new memories.

Synaptic Tagging and Capture

According to our hypothesis, retroactive interference occurs when newly encoded memories usurp resources that might otherwise be used to consolidate previously encoded memories. Under some conditions, however, a new memory may have the opposite effect of providing consolidation resources to weak memories[114–116]. The synaptic tagging and capture hypothesis posits that early-LTP produced by weak stimulation can be converted into late-LTP when a strong stimulus is delivered either before or after the early-LTP (i.e., the consolidation of weak Event A would be enhanced if preceded or followed by strong Event B)[117]. The analog to this is behavioral tagging and capture, where a short-term memory (e.g. a weak fear conditioning paradigm which lasts a few hours) can be converted into a long-term memory (e.g. 24hrs) if preceded or followed by a long-term memory[118, 119]. The idea is that a weak memory alone is insufficient to induce the molecular cascades leading to cellular memory consolidation. But, when another strong memory elicits such molecular cascades, it allows the weak memory to share resources (i.e. plasticity-related proteins) before it decays, thereby allowing it to consolidate into a long-term memory.

Interestingly, this intriguing result stands in apparent contrast to many other results in animals [43, 44] and humans [102, 120–122] showing that interference is often observed under conditions much like the ones that enhance the durability of weak memory in tagging and capture studies. Thus, considered in that light, the tagging and capture hypothesis may not always readily explain all phenomenon. The time interval between the two tasks may be a critical variable, and thus, will be an important issue to address in future studies.

Conclusion

In summary, we propose that it is not a coincidence that the mechanisms of cellular and systems consolidation are favored during periods of reduced encoding and hippocampal plasticity (including, but not limited to, SWS). According to the Opportunistic Consolidation Hypothesis that we put forth here, both phases of consolidation are compromised when hippocampal encoding mechanisms are engaged (i.e., when new memories are being formed). Furthermore, when hippocampal encoding mechanisms are disengaged without compromising hippocampal consolidation mechanisms, the consolidation mechanisms are released (Figure 3). We hypothesize that cellular and systems consolidation may be closely connected processes that unfold together in the hours after learning (with systems consolidation continuing for an extended period of time thereafter), rather than considering them as processes that unfold independently of each other on vastly different time scales. If so, this might explain why the conditions that favor one form of consolidation also appear to favor the other (Figure 3). According to the Opportunistic Consolidation Hypothesis, these potentially related consolidation processes preferentially unfold in an opportunistic fashion whenever the hippocampus is disengaged from the process of encoding new memories. Moreover, according to this view, the processes of declarative memory consolidation do not strengthen memories such that performance is enhanced from a baseline (Figure 1). Instead, as forgetting occurs (i.e., as memories weaken), both cellular and systems consolidation serve to render those memories more resistant to interference caused by the encoding of subsequent memories. These considerations suggest that for declarative memory, the issues of consolidation and interference are intimately connected, although future studies are needed to directly address this (Box 1).

A reduction in interference may also play a key role in the consolidation of non-declarative memories. However, in that case, the behavioral phenomenon to be explained (i.e., an absolute enhancement in performance, Figure 1) is different, suggesting different consolidation mechanisms as well. We assume that non-declarative memory benefits when the cortex is released from input – input that comes from sensory processing during the awake state or from hippocampus-to-cortex neural replay during SWS. Under these conditions of reduced input, combined with a cortically activated state relative to SWS and in some cases waking[123], cortical circuits may be free to extrapolate, which may account for the absolute enhancement of transitive inference[11], creativity[8], and perceptual learning[6] as a function of REM sleep. However, the story seems quite different for declarative memory. In our view, it is not sleep, per se, that is essential for the systems consolidation of declarative memory. Instead, we suggest that cellular and systems consolidation are interlinked, opportunistic processes that capitalize on periods of reduced encoding to render recently learned memories more resistant to interference.