The brain in motion: How ensemble fluidity drives memory-updating and flexibility

1Memories are neural patterns that guide behavior in familiar situations by preserving relevant information about the past. While this definition is simple in theory, in practice, environments are dynamic and probabilistic, leaving the brain with the difficult task of shaping memory representations to address this challenge. Dynamic environments imply that whatever is learned from a single episode may not hold true for future related experiences and should therefore be updated over time. If not, memory systems will fail to generalize to future retrieval episodes in which conditions may have changed, leading to suboptimal behaviors. Hence, the ‘goal’ of a memory system is not to remember individual events with the greatest possible precision, but rather to continually adapt its contents in order to build more and more accurate models of the world. To do this, the brain employs ‘memory-updating’, which we define as the process of modifying existing firing patterns to support the integration of new information into previously learned memories. Consistent with this concept, studies on reconsolidation have previously described how memories can become labile during retrieval allowing for memory-updating and these memory modifications continue indefinitely over an animal’s lifetime. While reconsolidation studies have greatly contributed to our understanding of continual learning, these studies typically rely on amnesic pharmacological agents that offer limited insight to how memory modification occurs at the neurophysiological and population level. However, recent studies utilizing approaches to observe and manipulate large-scale neuronal populations have reinvigorated the search for the flexibility of memory traces (Box 1). Indeed, as we will describe in this review, these observations support the concept of continual memory-updating and complement the findings regarding molecular mechanisms in the reconsolidation literature. Rather than discrete and fixed neural representations, we propose that memories are stored in flexible activation patterns that are continuously modified over experience, on the order of minutes to lifetimes.

2There is consensus that, generally speaking, memories are stored in activity patterns and synaptic weights of neuronal ensembles brain-wide, and that these ensembles are reactivated when recalling the memory. These ensembles can persist over long timespans measured over days to weeks, making them attractive substrates for long-term memory storage. The foundation for these ideas came from Donald Hebb, whose theories on synaptic plasticity and the stabilization of cell ensembles laid the groundwork for contemporary ideas in memory representations within neuronal networks. For instance, demonstrations of ensemble stability were found in hippocampal place cells (pyramidal neurons that fire according to the animal’s position in space) that were found to be stable over many weeks. Stability of neuronal ensembles was also supported by studies using localization of immediate-early gene expression in the hippocampus—exploration of two identical environments 20 min apart induced activity-dependent Arc expression in highly overlapping populations of CA1 neurons. In the amygdala, reactivation of a neuronal ensemble active during learning was correlated with memory recall several days later, indicating a stable neural correlate for fear memory. Based on these and related studies, modern theories suggest that dedicated populations of neurons (‘engram cells’) encode and store memories in the manner of a Hebbian cell ensemble.

3While groundbreaking, the discovery of stable ensembles as substrates for memories is an incomplete account of how memory systems operate over the course of an animal’s lifetime. Above all, these principles do not explain how the brain can integrate new experiences with old memories. In practice, some degree of flexibility must complement persistence in the successful implementation of memory. The ‘stability-plasticity dilemma’ describes the necessary compromise between these two opposing forces, allowing new learning to occur while preserving existing knowledge. Indeed, modeling studies have shown that an overly rigid neural network actually encumbers the acquisition of new information. In an inflexible network, existing knowledge can interfere with the encoding of new information (proactive interference) and is also subject to erasure during new learning (catastrophic forgetting). Therefore, in addition to understanding how memories can persist in neural populations, it is equally important to understand how memory systems can overcome collisions between old and new memories. To that end, dynamic memory ensembles encapsulate how memories can be both persistent and fluid. This idea is in line with the research on reversal learning, reconsolidation, schema learning, and systems consolidation, which all describe how previously learned behaviors can be modified to accommodate new learning. In particular, prominent theories on systems consolidation stress the importance of both persistence and flexibility—the hippocampus is often thought of as the flexible learner that trains neocortical networks to store memories long term. However, neocortical networks still undergo continual modifications as the animal learns over a lifetime.

4Although much is known about the flexibility of behaviors, we know much less about how the memories enabling those behaviors are updated at the neurophysiological level. While the flexibility of memory has been well appreciated in synaptic neurobiology and cognitive psychology, it has been largely neglected by neurophysiologists with a few recent notable exceptions. We begin by highlighting recent longitudinal observations of slow fluctuations in neural activity and synaptic structure that complement the ‘stable engram’ hypothesis. Intrinsic fluctuations that alter synaptic connectivity and cellular excitability provide an ever-present reservoir of flexibility in population activity patterns to store a memory. We propose that these slow fluctuations act like a conveyor belt that continuously supplies potential new storage sites of future memories, as suggested by previous studies on 'memory allocation' and 'memory-linking'. Neurons encoding these future experiences may overlap with existing engrams, updating past memories. In the following section, we describe how ‘unstable’ ensembles are not in fact paradoxical, but instead are necessary for flexible memory and behavior. Then, we outline the steps of the memory-updating process which entails (1) a partial reactivation of a previously formed engram, (2) recruitment of neuronal populations into existing engrams based on their excitability and functional connectivity, (3) deployment of plasticity processes that modify these networks by integrating the new population, and (4) temporal coordination of neural activity within and across regions for brain-wide memory-updating.

5With the advent of functional imaging methods that can enable longitudinal tracking of single neurons over days to weeks, a picture is starting to emerge that memory representations are not as fixed as we might expect. Despite experimenters’ efforts to keep external environments consistent and with no observed changes in the animal’s behavior, the firing patterns of neuronal populations continue to evolve, a phenomenon known as ‘representational drift’ (for reviews see;; ). Repeated exposures to the same conditions produce neural representations of these highly familiar experiences that nonetheless fluctuate on the order of hours to weeks, even with stereotypical behavior. For instance, hippocampal ensemble activity deviates over time, as measured by population vector correlations to a reference point, despite no deviations in the behavioral task (Figure 1a and b;;;;;;; ). Similar fluctuations have been reported in barrel cortex, motor regions, lateral entorhinal cortex, and parietal cortex, suggesting that while all these regions differentially contribute to memory, drift may be a ubiquitous feature of neural systems that seems to threaten memory stability. Alternatively, drift could actually reflect the inherent flexibility of the neural code and stem from numerous parallel neurobiological processes including spontaneous synaptic remodeling, and dynamic changes in cellular excitability (Figure 1c and d;; ). Next, we will consider how these processes could potentially contribute to the way in whichdynamic neural codes can support memory flexibility.

6Synapses are constantly being remodeled, which could be one underlying factor for drift in ensemble codes. While neural activity is a well-known mediator of synaptic plasticity, the reverse relationship is also possible—synaptic structure and weights could influence neural activity patterns. Along those lines, when action potentials are blocked in cell cultures, synapses continue to turn over across days, demonstrating that spontaneous synaptic remodeling occurs even in the absence of neuronal firing, and that these intrinsic remodeling events could impact the participation of individual neurons within ensembles. Hippocampal dendritic spines have been found to completely turn over within weeks, presumably contributing to fluctuations in neuronal firing patterns via rearrangement of synaptic weights. Supporting this possibility, one experiment described how local activity patterns at CA1 dendrites predicted longitudinal place field stability based on somatic calcium activity. The experimenters found that place cells were more likely to lose their field if the dendritic activity of those cells was more variable on a trial-by-trial basis. This suggests that the computations occurring in dendritic compartments could be disturbed by synaptic turnover, which in turn could potentially destabilize place fields across the whole population. Recent modeling efforts have shown that synaptic turnover can give rise to drifting ensembles. However, direct evidence for synaptic turnover as a major source for representational drift has not yet been shown and would require technically challenging feats of simultaneously imaging spines and somatic activity over long timescales.

7While seemingly disruptive, there is substantial evidence that constant flux does not preclude memory persistence. Although dendritic spines are routinely formed and eliminated, a significant proportion is estimated to persist over an animal’s lifetime, suggesting that long-term memories could be stored in the synapses located at those spines. These persistent spines are more common in the neocortex than in the hippocampus, consistent with systems consolidation theories suggesting that neocortical networks play a larger role in storing relatively more stable representations. Nonetheless, up to 15–25% of hippocampal CA1 neurons retain the same spatial firing patterns across weeks, which is sufficient for accurate spatial decoding. Such patterns may persist in part due to the relatively high resilience of larger dendritic spines, which has been observed in both hippocampal and neocortical spines. Additionally, modeling work has suggested that repeated offline reactivation of specific ensembles could maintain potentiated synaptic weights. Thus, long-term memories may be supported by a ‘backbone’ of stable spines and neurons that store gross features while the remainder might continually undergo plasticity to encode more detailed representations. A complementary ‘memory indexing’ theory has proposed that hippocampal neurons reinstate neocortical activity patterns for memory retrieval. In such a regime, neocortical storage sites may house relatively more stable memories that hippocampal computations incrementally modify.

8Even with the stability of a ‘backbone’ neural network, a nontrivial portion of neural networks is dynamic. Still, this does not appear to hinder memory or behavior. Performance on a variety of spatial navigation tasks remained stable across days despite continuous reorganization of firing patterns in parietal cortex and hippocampus. Most surprisingly, motor patterns are unchanged despite drift in neural activity from motor areas (;, but see ). Perhaps behavioral stability could be attributed to the consistency of the overall population regardless of the activity of individual neurons. Others have shown that from a population standpoint, the variability of any individual neuron is inconsequential to the fidelity of the overall neural code. A recent study proposed that downstream readers can hypothetically compensate for drift, suggesting that as long as the interpreter of an upstream neural code is capable of re-weighting its inputs, a stable readout can still be achieved. Although the neurobiological principles governing this re-weighting have yet to be determined, they may hinge upon neuromodulatory feedback signals such as dopamine or acetylcholine. Taken together, these studies indicate that memory retrieval can tolerate some degree of dynamism provided that some core backbone of the collective network remains intact or if downstream readers adjust their outputs in accordance with their fluctuating inputs.

9What is the function of drift in memory systems? We posit that drift can slowly and stochastically provide neural substrates that can bind new information, both for forming new memories and for updating old ones. In order to form new memories or update past memories, neural networks are faced with the formidable problem of having to potentiate the appropriate synaptic patterns among the huge number of possibilities that make up the synaptic connectivity space. Structural synaptic turnover through continuous spine degradation and formation could facilitate future learning by maximizing sampling across this synaptic connectivity space, increasing the likelihood of achieving certain spiking patterns, and ultimately potentiating their corresponding synaptic weights. Consistent with this logic, spine turnover is critical for birdsong acquisition, fear conditioning, and spatial navigation in zebra finches and mice. measured spine turnover in the mouse retrosplenial cortex and found that turnover rates, even before fear conditioning and spatial exploration, positively correlated with individual ability to learn each memory. In other words, high spine turnover rates provided a greater number of new spines available for memory encoding but may have also enabled faster sampling across synaptic space and therefore a quicker arrival to a synaptic connectivity pattern that adequately encoded the new information. Changes in synaptic connectivity could also heterogeneously influence the likelihood of spiking (intrinsic excitability) in neuronal subpopulations, which would in turn increase their likelihood of participating in future memory-encoding ensembles (i.e., memory allocation; Box 2;;;;; ). In this way, the brain could prioritize different rosters of neurons to diversify eligibility for memory encoding or updating. Such a framework is consistent with systems consolidation where information is constantly redistributed across the neocortical-hippocampal loop. In summary, spontaneous synaptic remodeling can supply additional synapses and neurons in which memories can be laid down as they are being experienced, contributing to memory flexibility.

10Drift may also facilitate the updating of previously learned memories by partially weakening old activity patterns in favor of strengthening new ones. In particular, hippocampal adult neurogenesis has been shown to play an important role in synaptic and circuit remodeling, supporting memory-updating and flexibility. The addition of newborn neurons in the dentate gyrus reconfigures existing hippocampal activity by decaying previously potentiated synapses. At first glance, this may appear disruptive for memory, but integration of these newborn neurons into hippocampal circuits can be useful because network-wide synaptic reorganization can erase obsolete patterns while making room for new ones. This feature is especially advantageous for memory-updating, which requires circuit reconfiguration in order for previous memories to accommodate potentially conflicting new information. In alignment with this hypothesis, exercise and environmental enrichment (which both increase neurogenesis) enhanced navigational flexibility when a goal was relocated. In these studies, mice with increased neurogenesis were better able to learn by modifying their expectations of where the goal was likely to be, but this ability was abolished when neurogenesis was inhibited. Importantly, neurogenesis was not required for initial memory acquisition, but only when the mice had to reverse their behavior. This suggests that neurogenesis plays an important role in memory-updating, and not just in acquisition of a new memory. Taken together, these findings suggest that after integration of adult-born neurons, the consequent synaptic reorganization allows experienced brains to acquire yet more information to integrate with previously learned associations.

11Aside from neurogenesis, spontaneous synaptic remodeling is a critical feature for permitting memory-updating through the exploration of synaptic connectivity space using prior knowledge as a starting point. Modeling studies have demonstrated that continual reconfiguration of synaptic weights is necessary for a network to adapt once dynamic conditions are imposed. One particular study used a neural network to perform a simulated motor task where specific outputs were rewarded. The authors then changed the output of a subset of model neurons relative to the motor task, thereby perturbing the neural network’s ability to produce the correct motor actions. However, because the neural network was able to continuously explore alternative solutions through a drift-like synaptic rewiring mechanism, it quickly adapted to the perturbation and regained high accuracy. This suggests that slow synaptic turnover may facilitate the ability to draw from a prior knowledge base (by storing connectivity patterns that are slow to decay) while still flexibly exploring related options through stochastic probing of new potential connectivity patterns, built atop existing ones. Such an implementation may underlie flexible behaviors that are based on memories for past outcomes. In other words, as long as a memory is retrievable and the ensemble encoding that memory is sufficiently plastic, it can be updated; but suppressing retrieval or plasticity blocks memory-updating. The notion of memory modification after retrieval has also been previously discussed in the context of reconsolidation, when memories are subject to modification after they are retrieved. In the next section, we examine how the brain modifies these memories.

12There is a large body of evidence showing that memories are not created de novo and instead draw upon prior knowledge stored in biophysical configurations of synapses and population activity. As such, memory-updating relies on instantiating an internal representation of past events, comparing expectations to current sensory input, and modifying the representation to improve predictive power. For example, in humans, viewing specific images reactivates the firing patterns observed during initial presentation of the stimulus. In rodents, similar neuronal populations are reactivated upon re-exposures to familiar environments, and activation of neuronal populations previously activated during the initial learning will induce memory retrieval.

13Reactivation of neural patterns during memory retrieval may constrain the brain to these state spaces, restricting the degrees of freedom for exploring possible solutions for new encoding. These states may act as continuous attractors and function as workspaces for memory-updating within that time window (Figure 2). Such a mechanism would imply that learning something associated with a previously learned stimulus would reactivate that previous memory. This view is again consistent with reconsolidation experiments where a reminder cue reinstates a previous memory, creating an opportunity for modification of that memory. Functional imaging experiments in the human brain found that presentations of reminders triggered reactivation of hippocampal and prefrontal networks, which facilitated learning of new, related items. Rodent studies have also found reactivation of hippocampal ensembles during memory-updating. Previous training on memory tasks with one set of stimuli accelerated their ability to form new associations with different stimuli. Upon introduction of these new stimuli, neurons that encoded previous, related stimuli were reactivated, suggesting that this population acted as a scaffold upon which new, related stimuli were embedded. This reactivation facilitates the connection of related experiences across time, allowing for the integration of events through the co-activation of neurons encoding past and present experiences. Such a mechanism is consistent with reports of optogenetic stimulation to co-activate ensembles to update previous memories with experimenter-defined ‘false memories’.

14To determine which experiences get integrated into a past memory, the relevance of input patterns might be weighted by temporal proximity to when the memory was encoded—events that occur close in time can become linked within neural networks. A functional imaging study in humans found that over two encoding episodes, subjects were more likely to combine their memories of overlapping items (object pairs) if they occurred within 30 min of each other. Within encoding episodes (defined by minutes-long temporal blocks), there is high similarity in hippocampal activity patterns in both humans and rats, suggesting that the network resides in the same state space during these episodes. Likewise in mice, two contextual experiences within 5 hr of each other became represented by a common set of hippocampal CA1 neurons, but not when they were separated by 7 days. Furthermore, not only were those two experiences represented by an overlapping ensemble, but when one of those contexts was subsequently paired with shock, animals transferred the fear from the shocked context to the neutral context. This indicates that the two memories became linked, which was sufficient to update a recent neutral memory with an aversive association. In another study, mice were fear conditioned with two different tones, separated by 6 hr, and these two tone-shock associations were represented in the amygdala by a common set of neurons. When one of the tones was repeatedly presented without shock (a paradigm known to lead to extinction of freezing), it also reduced freezing for the other tone, suggesting that the two memories became linked within those 6 hr. These effects were not seen when longer time intervals separated the two encoding events. Yet another study was able to link a conditioned taste aversion to a fear conditioned response through an overlapping population encoding the two conditioned stimuli. These ‘temporal memory-linking’ studies demonstrate that events occurring in close temporal proximity can mutually impact and modify memory representations of surrounding events.

15During memory-updating, how does the network determine which neurons are integrated into an existing memory? Accumulating evidence indicates that excitability and functional connectivity influence the allocation of neurons into memory ensembles. This theory, known as the ‘memory allocation hypothesis’, suggests that the excitability of a neuron predisposes it for encoding a memory (Box 2;; ). Moreover, the excitability and connectivity of neuronal populations are constantly changing, which would suggest that different sets of neurons encode or update memories as these neurons increase in prominence from the perspective of the network. In other words, drift, as a result of constantly fluctuating excitability, synapses, and intracellular proteomes, define the degree to which single neurons compete to participate in encoding a memory based on their excitability and connectivity to an existing memory engram.

16During memory-updating, in order to homeostatically maintain ensemble sizes, other neurons must also decrease their roles in memory encoding, ‘exiting’ the ensemble by reducing their contribution to the memory representation (Figure 1c). Because synaptic potentiation is saturated in engram neurons, a neuron’s exit from an ensemble might reflect depotentiation, reducing its likelihood of being co-activated with the remainder of the ensemble and potentially allowing it to encode other information. Depotentiation could result from a new competing ensemble suppressing the activity of neurons from a previous ensemble. For example, ensembles formed during extinction may be inhibiting ensembles formed during fear learning. As secondary evidence, optogenetic inhibition of an engram results in recruitment of an alternative engram, implying that the ensemble that might have been ‘next in line’ to compete with the first acts as a fail-safe to encode the present experience, ensuring functional homeostasis. Such competition may also underlie the loss of firing selectivity (e.g., receptive fields) over time in some neurons, decreasing their contribution to reliable memory encoding of the original memory. The gradual departure of those neurons may allow a transition to another set of neurons that surface with new information to integrate into the existing memory.

17During memory-updating, memory allocation on the basis of neuronal excitability promotes the overlap between ensembles encoding memories close in time. High excitability biases neurons encoding one memory to be recruited to encode another event, thereby linking the two memories through a shared neuronal ensemble (‘allocate-to-link’ hypothesis). But how could associations be made between experiences that are more temporally distant? One possibility is through ensemble reactivation during memory retrieval, which prepares an old ensemble to integrate new ensembles. Because memory retrieval involves a reactivation of a neuronal ensemble, excitation of those cells could trigger a physiological state that primes them for integrating the activity patterns of new neurons into their network. In line with this idea, learning increases neuronal excitability. Early studies found that eyeblink conditioning and operant conditioning increased neuronal excitability in CA1 and piriform cortex by decreasing the magnitude of slow after-hyperpolarizations (due to decreases in the calcium-dependent hyperpolarization current that follows spike bursts). Additionally, recent experiments have shown that during encoding and retrieval episodes, excitability is elevated in specific populations—the engram cells. By reactivating these engram cells during memory retrieval, they become re-excited along with newly excited cells at the next encoding event, which could induce synaptic potentiation between these ensembles. In doing so, this would combine memories that are related in content (yet encoded distant in time) in a common set of neurons. Simply put, the brain could recruit neurons encoding new information to fire alongside engram cells into an ensemble that now incorporates elements of new and old firing patterns. This modified ensemble then represents the updated memory.

18While excitability could determine the identity of the neurons to be recruited into an ensemble during memory-updating, pre-existing firing patterns (i.e., functional connectivity) at the time of encoding could mediate the temporal structure of the new neural motif destined for eventual potentiation. The temporal structure of spiking patterns could then be interpreted by downstream reader regions. An observed phenomenon that supports the hypothesis of pre-existing temporal patterns is hippocampal ‘preplay’ (Figure 3a). Preplay refers to an ensemble of CA1 neurons that produces reliable sequential activity prior to spatial learning, which is then recapitulated by firing in the same order while exploring physical space (;, but see ). This finding again implies that memory representations are not created de novo and instead that pre-existing functional connections and premature synaptic patterns can become potentiated by experience (Figure 3b). That is, there is always likely to be a neuronal sequence that occurs more often than others by virtue of the current state of synaptic connectivity, and this sequence becomes most likely to encode the next upcoming memory. In this way, the current state of the network as it drifts is connected to ongoing experience through the storage of information in neuronal ensembles with precise temporal coordination.

19Neural sequences derived from preplay templates may undergo cyclical refinement, consistent with our views on continuous memory-updating and how pre-existing temporal structure can shape the patterns that emerge. Once a preplay pattern is potentiated, increasing its occurrence rate, it is considered to be ‘replay’ and it is likely that replay patterns could then take on the role of preplay templates to integrate upcoming, related events. Modification of these sequence templates would therefore be an efficient mechanism for flexibly learning new information that is consistent with pre-established networks, without the need for widespread restructuring. In support of this idea, pre-existing sequences are relatively preserved after learning but reorganization is still apparent; new neurons are added to a backbone sequence after spatial exploration (; Figure 3c). One study investigated the properties of neurons that begin to co-fire with an established sequence after spatial exploration of a novel environment. This study found that newly recruited neurons contained higher spatial information, suggesting that the neurons that are most informative about spatial regularities in the environment are preferentially added to specific positions along the sequence. Moreover, consistent with the memory allocation theory, the neurons that were recruited tended to have higher firing rates than those that were not. Finally, the recruitment of neurons is dependent on the degree of temporal coordination between those neurons both prior to and during experience, demonstrating that functional connectivity among the ensemble participants is critical for recruitment. Thus, both excitability and functional connectivity bias the composition and temporal coordination of recruited neurons during memory-updating. But to actually induce physical changes in the synaptic strengths of these ensembles, the brain must modify synapses using plasticity proteins in concert with temporally coincident neural activity.

20How do slowly drifting synaptic configurations and spiking patterns actually stabilize to store information? One major factor is the upregulation of experience-dependent plasticity, which can temporarily stabilize memory ensembles. Experience-dependent plasticity can then potentiate activity patterns of memory ensembles to allow for those patterns to be retrieved for the next time they are to be updated. Indeed, there is enhanced expression of plasticity-related proteins such as cAMP responsive element binding protein (CREB) in numerous brain regions following memory retrieval and updating. Artificial memory retrieval, through activation of engram cells, triggers plasticity in those cells and memory modification at the behavioral level. For instance, after fear-conditioning, chemogenetically activating CREB-expressing amygdala engram neurons induces freezing and also triggers reconsolidation-associated protein signaling cascades that may prime memory-updating. Blocking plasticity with protein synthesis inhibition prevents memory-updating and impairs learning-associated changes in firing patterns and synaptic turnover. The myriad protein signal pathways triggered by learning also diversifyneuronal electrophysiological properties, such as excitation and inhibition patterns. These signaling cascades may enable complex interplay among a heterogeneous population to support and balance different aspects of memory, such as stability and flexibility. For example, a recent study found that Fos- and Npas4-expressing granule cell ensembles in the dentate gyrus were respectively responsible for generalization and discrimination of contextual fear memory. Further study is required to fully understand how constellations of molecular components interact in the cellular and network milieu to support high order cognition.

21In addition to the mobilization of plasticity mechanisms, neurons also need to spike in a temporally coordinated fashion in order to be recruited. Once neurons become coordinated at the millisecond scale, the resultant ensembles can then persist over longer timescales owing to synaptic potentiation. Oscillatory patterns in the brain represent an organizational framework for grouping the spike timing of neuronal subpopulations, which has led to a large body of literature on how rhythms could subserve experience-dependent plasticity, and therefore also memory-updating. In the hippocampus, plasticity is thought to be mediated by temporally coordinated neuronal ensembles within brief time windows such as the sharp-wave ripple (SPW-R) envelope and individual cycles of the 4–12 Hz theta rhythm. Both theta oscillations and SPW-Rs are powerful instigators of plasticity and long-term potentiation (LTP). In vitro, theta-paced stimulation is the optimal frequency for inducing LTP in hippocampal slices. Importantly, the theta phase is critical for determining whether LTP or synaptic depression occurs, meaning that not only is the magnitude of theta important, but also the spike timing of neuronal ensembles within theta cycles. Recent studies have shown that CA1 neuronal spiking needs to be temporally organized within theta cycles in order to be recruited into SPW-R events after learning. Thus, theta oscillations corral ensembles for potentiation and memory-updating.

22After temporal coordination, SPW-Rs may accelerate the rate of plasticity among the newly recruited neurons during memory-updating to integrate them into the ensemble. Artificial induction of SPW-Rs strengthens Schaffer collateral synapses in vitro, but curiously also causes long-term depression in synapses that were not recently active. This result suggests that recently activated ensembles are preferentially strengthened while connectivity with less relevant neurons is eroded, resulting in enhanced signal-to-noise. Consistent with this view, hippocampal SPW-Rs refine and stabilize place field maps in animals learning novel environments. This process is facilitated by upregulating SPW-R prevalence during learning. As a rat learns a spatial memory task, SPW-Rs increase in duration, and optogenetically prolonging SPW-Rs increases behavioral performance and recruits neurons that have spatial fields in behaviorally relevant locations. This preferential recruitment implies a mechanism for prioritizing integration of neurons that encode pertinent information and consequently serve important functional roles. SPW-Rs likely bind old and new memories by co-activating the appropriate ensembles. The end product is an updated ensemble whose spiking patterns convey updated information about recently learned stimuli.

23Hippocampal SPW-Rs could also recruit extrahippocampal neurons, which might help transmit and integrate recent hippocampal computations with knowledge stored in the neocortex to facilitate brain-wide memory-updating. The reorganization of neocortical spiking patterns could then readjust local (cortical) synaptic weights in a way that promotes meaningful activation patterns, synthesizing updated memories that may support schemas. This cross-regional dialogue is a signature of systems consolidation, where the hippocampus is thought to train neocortical ensembles by developing and reinforcing population patterns over a lifetime of experiences. Indeed, widespread activation of cortical regions is time-locked to hippocampal SPW-Rs, and SPW-R coupling between neocortex and hippocampus becomes upregulated after learning. After learning, hippocampal SPW-Rs trigger spiking in downstream cortical ensembles, and also couple with cortical oscillatory events such as ripples, spindles, and delta waves. These neocortical oscillatory events might themselves facilitate local plasticity. Although the individual functions of each frequency band remain an active area of research, generally, cross-regional coherence triggers spike pattern reorganization in downstream readers of hippocampal SPW-Rs that could be the neurophysiological readout of memory-updating. Recent evidence suggests that SPW-Rs can route specific content, suggesting that different memories can be individually communicated to downstream readers and updated separately. Hippocampal sequences, and their downstream neo- and subcortical readers, could reflect a modified ensemble structure that balances both parsimony and newly acquired information after undergoing experience-dependent plasticity.

24In this review, we have discussed forms of memory-updating and their neurophysiological signatures. We have proposed that these processes can be understood through the modification of neuronal ensembles whose membership is determined by both existing circuitry between neurons and intrinsicdynamics within neurons. Our framework is based on the idea that the integration of ensemble activity creates functional affinities (i.e., related memories). The organization of the neurons within these ensembles is heavily influenced by a combination of cellular excitability and functional connectivity (pre-existing temporal activity patterns) and the enactment of plasticity that modifies synaptic weights. While substantial work has shown that ‘stable’ engram neurons underlie memory retrieval, they do not necessarily persist indefinitely without modification. Instead, to satisfy the stability–plasticity dilemma, ensemble activities across brain regions must necessarily be fluid in order to update past memories with new information.

25While our review discusses how engrams change over time and experience, there is much still unknown about how the brain resolves the end products of memory-updating. How do these new modifications reconcile with established patterns? Previous studies have found that learning induces enduring changes: population activity does not revert to its original state when a familiar rule is reintroduced. These findings suggest that interleaved learning episodes can have lasting effects on the computational outputs of mnemonic systems, reflecting acquisition of new knowledge. One particularly interesting idea is that this new knowledge might be integrated with the old via the recombination of sub-ensembles, facilitating the formation of new relational topologies.

26An interesting effect of dynamic ensembles is that ‘tuning’ from neuronal populations relative to the external world deteriorates over time, which explains many observations of apparent neural instability. That is, the patterns of neuronal responses related to external stimuli are short-lived. Instead, the internal mappings between neurons persist through time. From the reference frame of the environment, the neural code for some variable (e.g., spatial location) may be drifting, but from the reference frame of the neurons, representations might remain stable due to compensatory adaptation at the network level. Such an outcome may arise from slow, coherent representational shifts such that different brain regions collectively shift in a coordinated manner. Among these regions may be ‘readers’ that interpret the messages of memory ensembles, and if those networks remodel coherently with upstream inputs, this could support the continuity of memories over time while also providing a means to update those memories. Thus, it is important to consider how inter-cellular activity patterns are modified during learning, and one major future avenue of research should be to determine how memory-updating rearranges internal activity patterns in an organized manner that retains past information while integrating new data. Most importantly, this perspective forces us to re-assess our very notions of stability. Nearly all central nervous system neurons are many synapses away from direct contact with physical stimuli, so if the upstream populations that do supply their input are inherently dynamic, we should rarely expect a neuron to maintain a stable relationship to an external variable even in the context of memory. Instead, we take the view that brain dynamics help generate unique combinatorial patterns onto which both new and familiar experiences are embedded.

27While many outstanding questions persist (Box 3), emerging technologies will prove critical for enhancing our understanding of the dynamic brain. In particular, voltage imaging has the advantage of being able to track neurons longitudinally while also providing exquisite temporal resolution and sensitivity to signals that are undetectable with calcium imaging (e.g., subthreshold depolarizations). Combining this with other optical techniques in freely behaving animals will allow us to advance our understanding of the dynamic brain.

28- What is the neurobiological milieu that determines the rate of representational drift? Do neuromodulators, sleep, or neuroendocrine signals exert influence?

29- Are drift rates in brain regions related to their local intrinsic synaptic volatility? What are the consequences of these differences and how might this relate to the rate of memory-updating in hippocampal versus neocortical structures?

30- How does ensemble remodeling occur in order to retain stored information while adding new information?

31- Downstream of memory representations, what are the circuits and computations that transform these firing patterns into behavioral decisions?