Memory replay refers to the reactivation of neural activity patterns that were originally expressed during experience, typically occurring later during rest or sleep. In the mammalian brain, this phenomenon has been studied most extensively in the hippocampus, where ensembles of place cells fire in sequences that mirror previously experienced trajectories through an environment. During active behavior, individual place cells are tuned to specific spatial locations, creating a population code that represents the animal’s position. During subsequent quiet wakefulness or non-REM sleep, these same cells fire again in similar or compressed sequences, as if the brain were internally “revisiting” the past episode.
At the cellular level, memory replay depends on the interplay between intrinsic excitability, synaptic plasticity, and local circuit dynamics. Synaptic connections that were strengthened during behavior via mechanisms such as long-term potentiation bias the likelihood that particular neurons will fire together again. When spontaneous fluctuations in membrane potential or network activity push some neurons toward threshold, these plastic connections help reconstitute the previously established firing sequences. Short bursts of synchronized activity propagate along the strengthened pathways, effectively reconstructing fragments of earlier experiences in the absence of external input.
In the hippocampus, replay is tightly coupled to characteristic network oscillations. During non-REM sleep and periods of quiet wakefulness, the CA3 and CA1 subfields exhibit sharp-wave ripple events: brief episodes of high-frequency activity nested within a slower field potential deflection. Sharp waves are thought to originate in recurrent networks of CA3, where dense excitatory connectivity can spontaneously ignite coherent population bursts. These bursts then drive fast oscillatory ripples in CA1 and downstream targets. Within the window of a single ripple, time-compressed sequences of place cell firing unfold, recapitulating trajectories that took seconds to traverse during waking behavior in only tens to hundreds of milliseconds.
Time compression is a central feature of memory replay, enabling the rapid reexpression of extended experience. During behavior, sequences of place cell activity are coordinated by theta oscillations, with cells firing at specific phases as the animal traverses their preferred locations. In offline states, these theta-organized sequences are “collapsed” into brief sharp-wave ripple events. This compression accelerates the reactivation of entire episodes, facilitating efficient synaptic modification in target structures such as neocortex, striatum, and amygdala. The accelerated timescale also allows multiple episodes to be replayed repeatedly during a single sleep period, increasing opportunities for consolidation and integration across memories.
The hippocampal circuitry is particularly well suited to generate replay because of its architecture. CA3 contains recurrent excitatory collaterals that form an autoassociative network, capable of pattern completion and sequence generation. When a subset of neurons in CA3 is spontaneously activated, recurrent excitation can propagate through synaptic pathways shaped by past experience, reinstating full patterns or ordered chains of activity. CA1 serves as a readout and transformation stage, integrating CA3 input with signals from entorhinal cortex. This arrangement allows stored sequences to be reactivated internally while still being broadcast to widespread cortical and subcortical regions that can use these signals to update their own synaptic weights.
Mechanistically, replay is influenced not only by excitatory connections but also by a diverse population of inhibitory interneurons that sculpt network timing. Parvalbumin-positive basket cells, for example, fire in tight coordination with sharp-wave ripples, providing rhythmic inhibitory control that organizes the precise timing of pyramidal cell firing. This temporal structuring helps preserve the relative order of spikes within a replayed sequence, ensuring that downstream synapses experience a temporally meaningful pattern of activity. Other interneuron classes regulate the onset and termination of sharp-wave ripple events, gating the windows during which replay can occur and shaping the mixture of sequences that compete for expression.
Neuromodulatory systems further regulate the conditions under which replay emerges. Levels of acetylcholine are typically low during slow-wave sleep and quiet wakefulness, a state that favors internal processing over the encoding of new external information. This neuromodulatory context promotes recurrent dynamics and the expression of previously strengthened pathways, supporting replay. In contrast, higher acetylcholine levels during active exploration prioritize sensory input and suppress some forms of spontaneous reactivation, shifting the network toward encoding rather than retrieval. Dopamine and noradrenaline can bias which experiences are preferentially replayed, with more salient, novel, or reward-predictive episodes reappearing more frequently during offline periods.
Although replay is often described in terms of spatial navigation and place cells, similar mechanisms have been observed in other hippocampal and cortical circuits encoding nonspatial information. Neurons that represent odors, task rules, or sequences of actions can also be reactivated in temporally ordered patterns after the original experience. In motor and prefrontal cortices, for example, ensembles involved in specific action sequences fire again during rest in a manner consistent with replay of learned motor routines or decision policies. This suggests that the fundamental mechanism of experience-dependent sequence reactivation is a general property of distributed brain networks, not restricted to spatial coding.
On a synaptic level, spike-timing-dependent plasticity provides a substrate for building and later expressing replay sequences. During behavior, when neuron A tends to fire shortly before neuron B, the synapse from A to B is strengthened. Offline, small fluctuations or spontaneous spikes in A can then trigger B, and so on down the chain, reconstructing the original firing order. Sharp-wave ripple–associated bursts provide ideal conditions for such timing-based learning and reactivation, because high firing rates and precise spike timing amplify the effects of plasticity rules. Over repeated cycles of experience and offline activity, these processes gradually refine the connectivity patterns that support robust replay.
Functional connectivity during replay is not static; it can be reshaped by the replay events themselves. Each episode of reactivation reexposes synapses to patterns of correlated activity, potentially strengthening or weakening them further depending on the timing relationships. This feedback loop allows the system to emphasize certain trajectories, prune rarely used connections, and interleave information from related episodes. In this way, replay does not simply reproduce the past but actively sculpts the memory networks that will underlie future behavior and, in later sections, provide a substrate for more anticipatory forms of preplay and neural prediction.
Beyond the hippocampus and neocortex, subcortical structures participate in and are influenced by replay. Striatal neurons, for instance, exhibit reactivation patterns that align with hippocampal sequences, especially in tasks involving reward learning and decision-making. When hippocampal place cell sequences corresponding to paths toward reward locations replay, associated striatal ensembles can also be reengaged, reinforcing the linkage between spatial trajectories and action values. Thalamic nuclei, particularly those relaying to prefrontal and sensory cortices, synchronize with hippocampal ripples, potentially coordinating the redistribution of information across large-scale brain networks.
Multiple timescales of replay coexist in the same brain, reflecting different computational roles. Very fast, highly compressed events during sharp-wave ripples may serve to train downstream circuits, integrate recent experience into existing schemas, and update value estimates. Slower, more extended reactivations during lighter sleep or quiet rest can support conscious or semiconscious recall, mental rehearsal, or the chaining together of episodes into narratives. These various forms of replay rely on overlapping but distinct mechanisms, such as differences in neuromodulatory milieu, network excitability, and the balance of hippocampal versus cortical contributions.
The occurrence and content of replay are also shaped by behavioral context and recent experience. Periods of intense learning, novelty, or uncertainty tend to be followed by richer and more frequent replay episodes. Moreover, not all experiences are replayed equally; trajectories that have been rewarded, that resolve uncertainty, or that are behaviorally relevant are often overrepresented. This selectivity suggests that replay mechanisms are integrated with systems that monitor outcome, error, and salience, ensuring that limited offline processing time is invested in memories most likely to influence future choices and learning.
Current evidence indicates that replay is not a monolithic process but encompasses forward, reverse, and fragmented reactivations. Forward replay reexpresses sequences in the same order as experienced, whereas reverse replay runs them backward, often immediately after reward consumption or task completion. Reverse sequences may be especially important for credit assignment, propagating information about outcomes backward along the sequence of states or actions that led to them. Fragmented replay, where partial sequences are stitched together, can blend components of different episodes, potentially contributing to generalization and the construction of more abstract representations.
At the level of systems neuroscience, these neural mechanisms of replay position the hippocampus as a central hub for orchestrating offline processing. However, the influence of replay extends across the entire “bayesian brain” architecture, in which neural populations encode and update probabilistic representations of the world. By repeatedly reactivating and refining internal models, replay helps adjust synaptic priors in cortical circuits, tuning them to the statistical structure of recent experience. This tuning prepares the brain for future inference and decision-making, setting the stage for how replay-related mechanisms can evolve into more explicitly predictive preplay dynamics.
Evidence for predictive preplay in the brain
Evidence that neural circuits can express preplay, rather than merely replay, emerged from detailed recordings of hippocampal activity in rodents. In classic spatial navigation experiments, ensembles of place cells are monitored as an animal first explores a novel linear track or open field. Surprisingly, in some studies, ordered sequences of place cell firing resembling trajectories through the yet-unexplored environment are observed during quiet wakefulness or sleep before the animal has physically traversed those paths. These preconfigured sequences, which are later expressed during actual movement in the same or a closely related order, suggest that the hippocampus can generate structured patterns that anticipate, rather than simply recall, future experience.
Preplay was initially identified by comparing neural patterns recorded during rest periods that preceded exposure to a new maze with patterns recorded during subsequent exploration. Researchers found that spontaneous sharp-wave ripple events in the hippocampus contained compressed firing sequences that statistically matched the order of place cell activation during later behavior. Control analyses, such as circular shuffling of spike trains and comparison with surrogate datasets, demonstrated that these correspondences exceeded what would be expected from random coincidence or general tuning similarities. This systematic alignment between pre-experience and post-experience sequences provided strong evidence that the hippocampal network can express predictive templates of upcoming trajectories.
Further support for preplay comes from experiments involving multiple distinct environments or track configurations. When animals are sequentially introduced to different mazes, some of the same hippocampal neurons can participate in different sequences, depending on the specific layout and task demands. In certain cases, sequences expressed during rest before exposure to a new maze are later used preferentially in that particular environment, even when alternative sequences would have been anatomically feasible. This mapping of pre-existing temporal structures onto specific spatial contexts implies that the hippocampus contains a reservoir of latent sequences that can be flexibly assigned to novel experiences, effectively serving as priors over possible future trajectories.
Importantly, preplay is not limited to one-dimensional paths. In two-dimensional open fields and complex mazes, analyses of place cell ensembles have revealed preconfigured patterns that correspond to multiple potential routes or branches. During rest, sequences spanning different arms or subregions of a maze can be observed, some of which are later realized as actual behavioral paths while others remain unrealized possibilities. This pattern is consistent with a generative process in which the hippocampus samples from an internal model of the environment’s structure, producing candidate trajectories that may or may not be adopted in subsequent behavior. The coexistence of realized and unrealized preplayed sequences highlights that preplay reflects probabilistic prediction rather than deterministic foretelling.
Evidence for predictive preplay also appears in tasks that emphasize decision-making and planning. When an animal pauses at a choice point in a maze, hippocampal ensembles can rapidly express sequences that sweep ahead along multiple potential paths, a phenomenon often termed vicarious trial and error. Some of these forward-looking sequences have been observed even before the animal has acquired extensive experience with all possible options, suggesting that the network is internally simulating candidate future states using partially learned or inferred structure. The preferential expression of sequences that lead toward future reward locations, especially as learning progresses, indicates that preplay is shaped by value signals and supports goal-directed prediction.
Studies combining hippocampal recordings with measurements in prefrontal cortex and striatum further strengthen the case for preplay as a predictive process. During rest periods preceding critical task stages, coordinated activity patterns arise across these regions that foreshadow specific action sequences or choice strategies. For instance, prior to a change in task rules, prefrontal ensembles can exhibit patterns that later reappear during successful implementation of the new policy, and these patterns are often temporally aligned with hippocampal sequences that represent relevant future states. Such cross-structure correlations suggest that preplay in hippocampus helps broadcast candidate futures to decision-making circuits, which in turn evaluate and select among them.
Evidence for preplay-like dynamics extends beyond spatial navigation and place cells, appearing in tasks that involve sequences of stimuli, actions, or abstract states. In odor sequence learning, hippocampal and cortical neurons can show anticipatory firing patterns that reflect upcoming items in a learned sequence, even when sensory cues are ambiguous or partially missing. During rest prior to sequence exposure or prior to performance on a given trial type, compressed activation patterns resembling the full sequence are sometimes observed, consistent with internal rehearsal of future sensory-motor chains. These anticipatory patterns are often modulated by task context, expected reward, and recent performance, reinforcing their functional linkage to prediction and planning rather than passive reverberation.
Sleep studies also provide compelling evidence for predictive preplay. Prior to initial exposure to a novel environment or task, hippocampal sharp-wave ripples during slow-wave sleep can contain structured sequences that later align with experience-dependent patterns. Some work suggests that these pre-experience sequences are influenced by prior learning in different but related environments, implying that offline processing can recombine existing representations to anticipate the structure of upcoming tasks. When animals are trained across days on families of mazes that share statistical regularities, the preplay content before a new variant reflects these regularities, as if the brain is using past data to form a generative model that biases future encoding.
In humans, direct single-neuron evidence for preplay is more limited due to methodological constraints, but converging findings from intracranial recordings, scalp EEG, and fMRI support the presence of predictive sequence activation. During rest prior to navigation in virtual environments, hippocampal and medial temporal lobe activity patterns can exhibit multivoxel signatures that later reappear during movement along specific routes. Additionally, during tasks that require learning sequences of images or locations, pattern classification analyses reveal that neural states associated with later sequence elements can be detected before those elements are actually presented, particularly when participants have strong expectations or explicit plans. These anticipatory activations share key properties with rodent preplay, such as temporal compression and dependence on task structure.
Another line of evidence comes from studies of mental simulation, imagination, and prospective memory. When individuals are cued to think about future scenarios, hippocampal and default network regions reactivate patterns related to relevant locations, people, or episodes in an order that mirrors the imagined unfolding of events. In some paradigms, spontaneous pre-task rest shows transient activations that resemble upcoming trial types or goal states, even when participants are not explicitly instructed to plan. The degree to which these pre-task patterns match subsequent task-related patterns correlates with performance and learning speed, suggesting that internally generated preplay supports more accurate and efficient future behavior.
The temporal structure of predictive preplay often mirrors that of replay but with a crucial directional bias toward the future. Like replay, preplay sequences are typically highly compressed, unfolding within tens to hundreds of milliseconds during sharp-wave ripples or related transient events. However, whereas offline replay often proceeds from recently visited states toward past starting points, preplay tends to progress from current or initial states toward potential future states or goals. This directionality is especially prominent when animals or humans are about to initiate a trial, make a decision, or enter a new environment, highlighting that preplay is intimately tied to preparation and foresight.
Critically, researchers have taken care to rule out simpler explanations for preplay, such as static anatomical biases or generic firing correlations. Control analyses frequently test whether pre-experience sequences match future trajectories more strongly than they match arbitrary or shuffled trajectories, and whether the same neurons participate in different sequences across contexts in a way that reflects task-specific constraints. The consistent finding that pre-experience patterns preferentially align with behaviorally relevant future paths or sequences supports the interpretation that preplay is an expression of an internal predictive model rather than a mere epiphenomenon of connectivity.
Emerging evidence indicates that neuromodulatory states and behavioral variables strongly influence the occurrence and content of preplay. Periods of elevated motivation, uncertainty, or anticipated change in task demands are often associated with richer preplay episodes prior to critical decisions or learning phases. Dopaminergic signals related to expected value and novelty appear to bias which candidate trajectories are preplayed, favoring those that lead to potential reward or efficient information gain. Changes in cholinergic tone, which modulate the balance between encoding and retrieval in hippocampus, also alter the prevalence of preplay, suggesting that the brain can flexibly tune its predictive simulation processes to match current goals and environmental demands.
Taken together, convergent findings from rodent, nonhuman primate, and human studies indicate that preplay is a robust and functionally significant phenomenon. Across species, the hippocampus and its associated networks do not merely replay the past but also generate structured, prospective sequences that align with likely or valuable future experiences. These predictive sequences are probabilistic, value-sensitive, and context-dependent, aligning naturally with the idea of a bayesian brain that maintains and updates internal priors about future states of the world. In this framework, preplay provides a neural substrate for sampling from these priors, enabling the brain to explore possible futures before they occur and to shape subsequent learning and decision-making accordingly.
Transitions from replay to preplay dynamics
The transition from reactivation of past experience to generation of prospective sequences can be understood as a gradual shift in how circuits organize and exploit their existing connectivity. Replay relies on synaptic structures laid down by prior experience; preplay exploits those same structures, plus additional latent organization, to generate trajectories that have not yet occurred. In the hippocampus, recurrent networks in CA3 function as a reservoir of partially structured activity patterns. When driven by different initial conditions and neuromodulatory states, that reservoir can either settle into sequences closely matching recent episodes (replay) or into trajectories that extend beyond them, effectively sampling possible futures (preplay). The underlying substrate is the same; what changes is how the network is initialized, constrained, and read out.
One important axis of transition is the balance between externally driven and internally generated input. During behavior, strong entorhinal input anchors hippocampal dynamics to ongoing sensory information, constraining place cells to fire in tight correspondence with the animal’s location or task state. After behavior, when external drive is reduced and acetylcholine levels fall, recurrent CA3 collaterals and intrinsic excitability play a larger role. Under these conditions, networks previously tuned by recent experience are free to revisit past trajectories in compressed replay. As training proceeds and the internal model becomes more stable, the same recurrent machinery can begin to extend beyond strictly experienced paths, stitching together segments into novel routes that anticipate plausible future states.
Another key dimension is the degree of constraint imposed by recently acquired memories. Immediately after learning, the hippocampal system is highly biased toward recapitulating those specific episodes; replay is dominated by concrete trajectories and closely mirrors the temporal order of experience. With consolidation and repeated offline processing, the connectivity underpinning these episodes becomes more abstract and overlapping, especially as related experiences are integrated. This overlap allows the system to recombine fragments of different episodes into new combinations. At this stage, offline activity begins to show hybrid sequences that start as veridical replay but deviate into unexplored branches, reflecting a continuum from strict memory retrieval to generative prediction.
Within individual sharp-wave ripple events, the transition from replay to preplay can occur on sub-second timescales. Some sequences begin with a fragment of a recently traversed path—a canonical signature of replay—but then jump across gaps in physical or experiential space to states the animal has never directly visited. Analyses of such “branching” events show that the first portion of a ripple aligns best with past behavior, whereas the later portion aligns with routes the animal will only explore in subsequent trials. These hybrid events suggest that the hippocampal network seeds prospective simulations by anchoring them to reliable past trajectories, then letting dynamics explore beyond the boundaries of experience.
Forward and reverse sequences provide another window into transitional dynamics. Reverse replay, often observed immediately after reward consumption, propagates activity from goal locations back toward starting points, supporting credit assignment over recent actions. As learning stabilizes and the environment becomes familiar, additional forward-looking sequences begin to appear before trial onset, some of which anticipate paths to future rewards. Across sessions, the relative frequency of reverse replay can decrease while forward preplay increases, marking a functional shift: the system moves from retrospectively assigning value to past actions toward prospectively evaluating candidate future policies.
Decision points in space and task structure highlight moments where replay and preplay blend. When an animal pauses at a choice point, hippocampal ensembles can rapidly alternate between sequences that look like backward replay of the just-traveled path and sequences that sweep ahead along potential options. Early in training, the backward component dominates, seemingly evaluating how the animal arrived at its current state. With practice, the forward component grows stronger and more selective, especially toward arms or routes associated with higher reward probability. This evolution illustrates how the same network event—compressing state sequences within a ripple—can gradually reorient from retrospective evaluation to prospective planning.
From the perspective of a bayesian brain, the replay-to-preplay transition reflects a reweighting of priors and likelihoods encoded in synaptic patterns. Initially, the hippocampal network’s priors over trajectories are weak, and offline activity is heavily constrained by the most recently observed sequences. As more data accumulate, synaptic plasticity tunes the network so that certain trajectories become high-probability paths in the internal model, while others are suppressed. Offline events then become samples from this updated posterior over trajectories. When sampling is heavily conditioned on recent evidence, offline events resemble replay. As conditioning relaxes and sampling explores the higher-probability regions of the trajectory space more broadly, events increasingly resemble preplay, in which the brain generates plausible future sequences consistent with learned structure.
Neuromodulatory systems provide a powerful lever for steering the network along this continuum. Low cholinergic tone facilitates retrieval and recurrent processing, supporting both replay and preplay, but the presence or absence of dopaminergic signals related to expected reward and novelty biases which type dominates. After surprising or highly rewarding outcomes, dopaminergic bursts strengthen synapses along the just-experienced path, increasing the likelihood that upcoming ripples will reexpress that path in reverse or forward replay. As the task becomes predictable and reward is less surprising, dopaminergic modulation increasingly favors trajectories that lead to efficient exploration or information gain, shifting the content of offline events toward preplay of untried but promising routes.
Behavioral context also critically shapes transitions. Immediately following intense exploration of a new environment, replay events are largely constrained to actually traversed paths, often in reverse order and closely time-locked to the end of trials. As the environment becomes familiar, offline activity during intervening rest periods and even during inter-trial pauses incorporates more prospective sequences—paths to shortcuts, alternative routes to the same goal, or trajectories that bypass previously visited dead ends. Experiments in which maze configurations change mid-session show that when prior knowledge generalizes to the new layout, hippocampal sequences quickly shift from replay of the old map to preplay of routes consistent with the new structure, sometimes before the animal has physically sampled all relevant paths.
At the local circuit level, inhibitory interneurons and network oscillations help determine whether an event will favor replay or preplay. Patterns of interneuron firing can gate which subsets of pyramidal cells participate in a given ripple, effectively selecting a subspace of the network’s latent sequences. When inhibitory gating emphasizes neurons strongly tied to recent sensory experience, the resulting event tends to reflect straightforward replay. When gating allows recruitment of neurons that share more abstract relational coding—such as cells that respond to conceptual or topological similarity rather than specific locations—the same oscillatory window can support sequences that generalize across contexts, a hallmark of preplay-like prediction.
Interactions between hippocampus and cortical regions further drive the transformation. Early in learning, hippocampal replay trains neocortical circuits by broadcasting compressed sequences that neocortex initially treats as novel input. Over time, cortical networks develop their own representations of task structure, including predictive codes for expected sensory and reward outcomes. As these cortical predictions strengthen, they feed back to hippocampus via entorhinal cortex, biasing which hippocampal sequences are likely to be reactivated or generated. Under this bidirectional coupling, hippocampal events are no longer mere echoes of experience but become coordinated with cortical expectations, pushing the overall system toward preplay that reflects integrated, multi-level predictions.
The temporal evolution of offline dynamics across learning provides empirical evidence for this systems-level shift. Early recordings during novel task exposure show many more events that can be classified as pure replay: sequences align closely with just-experienced trajectories, and their direction is strongly tied to immediate reward or movement history. Across days, however, an increasing fraction of events cannot be mapped neatly onto past behavior. Instead, they represent recombinations of familiar segments, shortcuts never taken, or paths that only become behaviorally realized later. This gradual enrichment of the repertoire of offline sequences reflects how an initially memory-bound system transforms into one that actively simulates futures constrained, but not dictated, by past experience.
Transitions from replay to preplay dynamics are not strictly unidirectional or permanent; rather, the system flexibly toggles along this continuum according to current demands. During phases of rapid learning or environmental change, replay resurges, reemphasizing accurate recapitulation of new episodes to stabilize encoding and credit assignment. Once the environment stabilizes, preplay becomes more prominent, enabling efficient exploitation of accumulated knowledge through prospective planning. This flexibility suggests that replay and preplay are best seen as complementary operating modes of the same predictive machinery, with neural circuits adjusting the balance between them as the relative importance of memory stabilization versus forward-looking prediction fluctuates.
Importantly, the same formal mechanisms that enable transitions between replay and preplay at the level of trajectories may also govern prediction in more abstract cognitive domains. As hippocampal and cortical networks internalize regularities in sequences of stimuli, categories, or actions, offline events progressively shift from simple reinstatement of specific episodes toward simulation of novel sequences that adhere to learned rules or schemas. In these settings, early offline dynamics echo particular training trials, while later activity anticipates correct responses in untrained combinations, mirroring the shift from episodic replay to schema-guided preplay observed in spatial tasks. This parallel underscores that the replay-to-preplay transition is a general organizing principle of predictive neural computation rather than a special feature of spatial navigation.
Computational models of neural prediction
Computational models of neural prediction aim to formalize how biological circuits, such as the hippocampus and associated cortical regions, implement processes like replay and preplay. At their core, these models attempt to capture how synaptic connectivity, neuronal dynamics, and learning rules jointly give rise to the generation of sequences that extend from past experience into plausible futures. Many frameworks treat the hippocampal–cortical system as implementing a form of probabilistic inference, where internal states represent hypotheses about environmental structure and future trajectories. In this view, prediction corresponds to sampling or computing from an internal generative model, with replay and preplay emerging as different manifestations of this sampling process under varying constraints and initial conditions.
One influential class of models casts the hippocampal formation as a generative sequence model, akin to a recurrent neural network trained to predict future states from past inputs. These models often incorporate attractor dynamics in CA3, where recurrent collaterals support pattern completion and propagation of activity along learned trajectories. During behavior, the network is driven by external sensory inputs, mapping current observations to internal states that encode position or task context. Offline, when external drive is minimal, the same recurrent dynamics can be run in a “free mode,” allowing the network to traverse stored state transitions autonomously. Depending on how strongly the network is constrained by recently activated states and synaptic weights, these autonomous trajectories appear as either faithful replay of experienced sequences or preplay of novel, but statistically likely, paths.
From a probabilistic perspective aligned with the bayesian brain hypothesis, hippocampal and cortical circuits are thought to encode probability distributions over states, trajectories, and task structures. In this framework, priors represent expectations derived from long-term experience, while likelihoods capture how current sensory inputs relate to internal states. Replay can be understood as sampling from a posterior distribution conditioned strongly on recent episodes, effectively revisiting states with high posterior probability given past observations. Preplay, by contrast, corresponds to sampling more broadly from the prior-structured trajectory space, possibly with only weak conditioning on immediate past states, thereby generating candidate futures that are consistent with learned statistical regularities but not yet experienced.
Many computational accounts formalize these ideas using Markov decision processes (MDPs) and reinforcement learning. Here, states correspond to spatial locations, task configurations, or abstract representations; actions link states via transitions; and value functions estimate expected cumulative reward. Replay and preplay are modeled as internal sweeps through the state space, updating value estimates or policies without external movement. In model-based reinforcement learning, these sweeps correspond to “planning backups,” where simulated transitions are used to refine estimates. Hippocampal sequences are proposed to instantiate such backups, with reverse replay implementing backward propagation of value from goal to start states and forward preplay implementing evaluation of potential future paths before action selection.
Successor representation models provide a particularly elegant bridge between neural sequences and prediction. The successor representation encodes, for each state, the expected discounted occupancy of future states under a given policy. Computationally, it can be learned incrementally via temporal-difference updates during experience, but it also admits efficient updates via offline simulation. In neural implementations, hippocampal place cells or more abstract state cells are proposed to encode elements of this predictive map. Replay sequences re-sample known trajectories, refining the successor representation by reinforcing predictable state transitions. Preplay sequences, especially those that traverse unvisited shortcuts or novel combinations of known segments, update expectations about future state occupancy beyond directly experienced paths, enabling rapid generalization and one-shot adaptation.
Another family of models treats hippocampal networks as reservoirs or dynamic generative models that learn a low-dimensional manifold of environmental structure. In such models, experience shapes a connectivity matrix such that intrinsic dynamics naturally give rise to trajectories that mirror environmental topology. During encoding, ongoing inputs carve paths through this manifold, etching synaptic traces via spike-timing-dependent plasticity or Hebbian rules. Offline, noise or small perturbations can cause the network to re-enter these paths (replay) or explore nearby, untraveled branches of the manifold (preplay). The degree of exploration depends on parameters such as noise amplitude, gain, and adaptation, which can be mapped onto neuromodulatory states in biological circuits.
Generative models based on probabilistic graphical structures further elucidate how the brain might internalize and use relational knowledge for prediction. In these models, nodes represent states or events, and edges encode probabilistic dependencies or transitions. Learning adjusts edge strengths to capture co-occurrence and temporal order statistics. When such a graphical model is embedded in a neural network with recurrent connectivity, sampling algorithms like Gibbs sampling or particle filtering can be implemented through stochastic neural dynamics. Replay arises when the sampler is initialized near states corresponding to recent experiences and allowed to run forward or backward along high-probability edges. Preplay occurs when the sampler is initialized in more abstract or contextual states (such as a goal or task schema), and the network generates candidate sequences that start from or lead to those states, effectively simulating possible futures.
Grid–place cell models of hippocampal and entorhinal function provide concrete instantiations of predictive sequence generation in spatial domains. In these models, entorhinal grid cells form a metric representation of space, and their interactions with hippocampal place cells support path integration and localization. Predictive coding variants propose that the combined grid–place network not only represents current position but also encodes velocity and directional information, enabling extrapolation into the near future. Offline, the same circuitry can simulate movements along the internal spatial map by iteratively shifting grid phases and activating associated place cells, producing replay-like and preplay-like sweeps. Planning algorithms, such as A* or value iteration, have been mapped onto such neural architectures, with sequences of place-cell activation representing exploration of possible routes through the internal map.
Hierarchical models extend these ideas beyond simple trajectories to structured tasks and schemas. In hierarchical reinforcement learning and options frameworks, higher-level states represent subgoals or task chunks, and policies operate at multiple temporal scales. Computational accounts suggest that hippocampal sequences can reflect not only primitive state transitions but also transitions between subgoals, encoding temporally abstract “options.” Replay at this higher level consolidates successful strategies, while preplay simulates combinations of options that have not yet been executed, enabling flexible recombination and transfer. Neural implementations posit parallel sequence generation at multiple timescales, with fast, fine-grained place-cell sequences nested within slower, more abstract prefrontal or cortical sequences, together supporting multi-level prediction and planning.
Deep learning models, particularly recurrent and attention-based architectures, have been used to emulate hippocampal-like prediction in more complex and non-spatial domains. Sequence-to-sequence models, predictive coding networks, and transformer-based systems can be trained to anticipate future items in sequences of images, sounds, or abstract symbols. Once trained, these networks often exhibit internal activations that recapitulate experienced sequences during “offline” or generative phases, analogous to replay, and can generate novel but statistically coherent continuations, akin to preplay. By analyzing these artificial networks, researchers can probe how different learning rules, architectures, and regularization strategies affect the emergence of compressed, temporally extended internal sequences and compare these patterns against neurophysiological data.
Computational work has also addressed how neuromodulation shapes the balance between replay and preplay and, consequently, the character of prediction. Models incorporating global parameters that modulate plasticity rates, gain, or exploration–exploitation balance demonstrate how dopamine- or acetylcholine-like signals can shift networks between modes. For example, elevated dopamine can increase the probability of replaying reward-associated trajectories, enhancing value learning, while lower dopamine or increased uncertainty signals can encourage sampling trajectories that maximize information gain, leading to exploratory preplay of untried routes. By systematically varying these parameters, modelers can reproduce experimentally observed shifts in sequence content across learning stages and motivational states.
A crucial theoretical challenge is to reconcile the apparent discreteness of ripple-associated sequences with continuous-time control and perception. Some models address this by treating offline events as Monte Carlo samples that update value functions or predictive maps, while online behavior is governed by function approximators updated by these samples. In this scheme, replay and preplay provide training data for downstream controllers during pauses, sleep, or inter-trial intervals, enabling the system to improve policies without constant interaction with the environment. The frequency, content, and directionality of sequences determine which parts of the state–action space receive the most offline training, and learning algorithms can be designed so that this prioritization matches observed neural biases toward behaviorally relevant trajectories.
Computational approaches have also been extended to account for non-spatial and more abstract predictive cognition, consistent with evidence that the hippocampus supports relational and episodic prediction. In relational models, states represent combinations of features, concepts, or social entities, and transitions capture inferred relationships or narrative structures. Replay in such models supports memory consolidation by reinforcing co-occurrence patterns and temporal relations between events. Preplay simulates novel combinations of known elements that obey learned relational constraints—for example, imagining a new story consistent with familiar characters and settings. Formalizing these processes requires generalizing trajectory-based reinforcement learning algorithms to operate over graph-structured or high-dimensional latent spaces, a direction increasingly explored with graph neural networks and structured generative models.
Another domain where computational models of neural prediction intersect with data is in explaining variability and stochasticity in sequences. Empirically, replay and preplay are not perfectly deterministic; they exhibit trial-to-trial variability and generate many unrealized sequences. Sampling-based models account for this by explicitly representing uncertainty and treating sequences as stochastic draws from a probability distribution over trajectories. In such models, the variance and covariance structure of these distributions determine the diversity of generated sequences. High uncertainty leads to broad exploration in preplay, whereas low uncertainty promotes more convergent replay of specific high-probability paths. Adjusting these uncertainty parameters allows models to mimic the shift from exploratory preplay during early learning to more focused replay once tasks are mastered.
Computational frameworks have further begun to address how cortical networks integrate hippocampal sequences into long-term predictive models. One approach posits that neocortex implements a slower-learning, high-capacity predictive coding system that maintains structured generative models of the world. Hippocampal replay provides targeted “training samples” to cortex, emphasizing surprising, salient, or reward-relevant episodes. Over time, cortical generative models come to encode many of the same predictive relationships, to the point where cortical circuits can support prediction even when hippocampal contributions are reduced. Preplay in hippocampus then becomes a way of querying or testing these cortical models, generating hypothetical scenarios that are evaluated and refined via bidirectional interactions between the two systems.
Across these diverse modeling approaches, several common computational principles emerge. First, prediction is treated as a consequence of having a generative model that captures both the statistical regularities of past experience and the dynamics of state transitions. Second, replay and preplay appear as natural modes of operation of such models when run offline or under weak sensory drive, with replay corresponding to strongly conditioned inference and preplay to more exploratory sampling from structured priors. Third, learning rules that couple sequential activation with synaptic modification—for example, temporal-difference learning or spike-timing-dependent plasticity—provide a mechanism for internal sequences to sculpt the very models that generate them, yielding a closed loop of prediction, evaluation, and refinement. These principles furnish a computational scaffold for interpreting neural data on sequence generation and for designing future experiments that can discriminate among competing theories of how the brain implements predictive cognition.
Implications for cognition and behavior
The interplay between replay and preplay has far-reaching consequences for cognition by shaping how organisms use past experience to guide future behavior. Offline sequences in the hippocampus and associated networks do more than stabilize memories; they actively reorganize internal models in ways that alter perception, decision-making, and action selection. By repeatedly sampling from learned structures, these sequences tune the brain’s implicit priors about what is likely to happen next, biasing subsequent interpretation of ambiguous stimuli, the evaluation of options, and the initiation of goal-directed behavior. In this sense, the “bayesian brain” is continually recalibrated offline, such that waking cognition unfolds against a background of predictions refined during sleep and rest.
One major implication concerns planning and flexible navigation. Forward-looking sequences in place cells and related systems provide a neural substrate for mentally exploring alternative routes before committing to a path. When an animal faces a novel choice point, preplay sequences can sweep ahead along different branches, effectively evaluating potential outcomes without overt movement. This internal simulation enables rapid selection of efficient routes, including shortcuts that have never been physically traversed but can be inferred from the structure of the environment. Behavioral experiments show that animals can exploit such inferred paths after minimal direct experience, and the presence and quality of preplay sequences prior to these choices correlate with successful use of shortcuts and rapid adaptation to maze reconfigurations.
Decision-making beyond spatial domains is similarly influenced by predictive sequence activity. In tasks where subjects must select between actions with different rewards or risks, hippocampal–prefrontal interactions support deliberation by representing possible future states and their expected outcomes. Offline reactivation of action–outcome associations enriches internal models of the task, allowing individuals to update their policies even in the absence of immediate feedback. When the same circuits later express preplay at the moment of choice, they effectively “rehearse” the consequences of candidate actions, aligning behavior with long-term goals rather than immediate impulses. Disruptions to this machinery, whether through lesions, neuromodulatory imbalances, or sleep deprivation, often manifest as myopic decision-making, diminished flexibility, or increased susceptibility to maladaptive habits.
Memory consolidation and transformation constitute another central behavioral consequence. Replay during sleep strengthens associations within specific episodes, but it also promotes integration of related experiences into broader schemas. Over repeated offline cycles, memories become less tied to the precise circumstances of encoding and more organized around abstract regularities. This process supports generalization: individuals become able to apply learned rules to new contexts, infer missing information, or complete patterns they have not seen in full form. Preplay capitalizes on this abstraction by generating novel combinations of schema-consistent elements, effectively testing and refining generalized knowledge. Behaviorally, this manifests as insight, transfer learning, and the ability to solve problems that require going beyond literal past experiences.
Predictive sequence activity also plays a role in perception and anticipatory attention. When internal models have been shaped by extensive replay, sensory systems come to expect particular sequences of stimuli in given contexts. As a result, neural responses to expected inputs are sharpened and accelerated, while unexpected inputs elicit larger error signals and capture attention. Preplay-like anticipatory activations can preconfigure sensory and motor circuits just before predicted events, reducing reaction times and enhancing accuracy. For example, when individuals anticipate a particular visual or auditory sequence, pre-activation of relevant cortical patterns can be detected during the preparatory interval, and the degree of this pre-activation correlates with perceptual sensitivity and speed of response.
Emotional learning and regulation are also shaped by replay and preplay. Offline reactivation of episodes involving reward, threat, or social feedback reinforces associations between contexts and affective value, influencing future approach or avoidance behavior. Reverse replay after receiving reward can strengthen links between distal cues and positive outcomes, supporting the development of adaptive preferences. Conversely, replay of aversive events can maintain or amplify fear responses. Preplay enters when the system begins to simulate possible future emotional outcomes: imagining threatening or rewarding scenarios engages many of the same networks and sequence dynamics as actual experience. This capability underlies anticipatory anxiety, optimism, and various forms of prospection, whereby individuals mentally project themselves into future situations and adjust their behavior in the present based on those simulated feelings and outcomes.
Sleep-dependent changes in cognition provide a particularly clear window into the behavioral impact of offline sequences. Across a wide range of tasks—spatial navigation, motor skills, rule learning, and relational inference—performance typically improves after sleep, even without additional practice. These gains often correlate with the occurrence and quality of hippocampal sharp-wave ripples and associated replay events. Moreover, targeted memory reactivation paradigms, in which subtle cues associated with prior learning are presented during sleep, can bias which memories are replayed and thereby selectively enhance specific skills or knowledge. Such findings indicate that waking cognition is partially determined by which experiences the offline system chooses to rehearse and how those rehearsals shape predictive models used the following day.
The content of replay and preplay is not neutral; it is selectively biased toward experiences and trajectories that matter for future behavior. Rewarded paths, novel situations, and episodes with high uncertainty are preferentially reactivated, leading to disproportionate strengthening of their representations. This prioritization ensures that limited offline processing time is spent on memories with the greatest potential impact on future choices. Computationally, it resembles prioritized experience replay in artificial reinforcement learning, where high-error or high-value transitions are sampled more frequently to accelerate learning. Behaviorally, it means that individuals are especially likely to consolidate and anticipate situations that carry significant consequences, supporting efficient adaptation in complex and changing environments.
At the level of cognitive strategies, predictive sequence activity supports the construction of mental narratives and event models. Humans and other animals often organize experience into structured episodes with beginnings, middles, and ends, and they use these structures to anticipate what will happen next. Replay helps maintain coherent narratives by reactivating sequences of events in order, preserving temporal and causal relationships. Preplay extends these narratives into the future, enabling individuals to mentally simulate how current storylines might continue under different choices or external conditions. This narrative-based prediction guides everyday cognition, from mundane planning (such as organizing errands) to complex social reasoning (such as forecasting the reactions of others in hypothetical scenarios).
The same mechanisms that enable adaptive foresight can, when dysregulated, contribute to cognitive and psychiatric disorders. Excessive or biased replay of negative or threatening episodes may underlie persistent rumination and intrusive memories, as seen in depression and post-traumatic stress disorder. If preplay circuits preferentially simulate catastrophic futures, individuals may experience chronic anticipatory anxiety and avoidance, even in relatively safe environments. Conversely, weakened or disorganized predictive sequences may impair the ability to anticipate consequences, contributing to impulsivity and poor planning in conditions such as attention-deficit/hyperactivity disorder or certain forms of frontal lobe dysfunction. Understanding how replay and preplay are altered in these conditions can inform interventions aimed at recalibrating predictive circuitry, for example through targeted cognitive training, neuromodulation, or sleep-based therapies.
Social cognition is another area where predictive neural dynamics exert substantial influence. Interacting with others requires anticipating their actions, intentions, and reactions, often on the basis of limited information. Replay of past social encounters can refine models of others’ behavior, highlighting which cues were predictive of particular outcomes. Preplay of potential interactions allows individuals to mentally rehearse conversations, negotiations, or cooperative strategies, adjusting their approach before acting. Such anticipatory simulations engage hippocampus, medial prefrontal cortex, and other default network regions, and their richness is associated with greater social competence, empathy, and the ability to navigate complex group dynamics.
On a broader scale, the continuous operation of predictive sequence mechanisms shapes an individual’s sense of agency and temporal perspective. The capacity to mentally travel in time—to revisit the past and project into the future—relies on the same hippocampal circuitry that generates replay and preplay. When these processes function effectively, individuals experience a coherent link between past actions, present states, and future possibilities, supporting long-term planning, delayed gratification, and the pursuit of stable goals. When they are compromised, temporal horizons can shrink, making it difficult to connect present choices with distant outcomes. This shift has concrete behavioral implications, from financial decision-making and health behaviors to educational and career planning.
Learning efficiency and creativity are also shaped by how internal sequences sample the space of possibilities. Replay that closely tracks experienced episodes tends to stabilize existing knowledge, reducing forgetting and noise. Preplay that explores untried combinations or paths can generate novel ideas and strategies, some of which may later be tested in behavior. Creative problem-solving often involves drawing connections between disparate experiences and imagining unconventional solutions; these processes are supported by offline recombination of memory fragments into new configurations. The same sequence-generating circuits that once encoded straightforward trajectories through space can, under more abstract representational schemes, traverse conceptual landscapes, enabling leaps in understanding and innovation.
The balance between online and offline computation has implications for how organisms allocate cognitive resources. By shifting some forms of learning, evaluation, and planning into rest and sleep, the brain frees online processing capacity for real-time perception and action. Replay consolidates what has been learned; preplay prepares the system for upcoming challenges; together, they reduce the need for extensive trial-and-error exploration in the moment. This division of labor helps explain why adequate sleep and quiet wakefulness are so critical for performance across domains and why chronic disruption of these states can erode not only memory but also foresight, emotional regulation, and adaptive decision-making. In everyday life, much of what appears as spontaneous insight, sudden understanding, or “gut feeling” at a decision point can be traced to prior offline sequences that have silently reshaped the predictive architecture of the brain.
