In physics, the notion of a thermodynamic arrow of time arises from the simple observation that many macroscopic processes are irreversible. A drop of ink disperses in water, but we never see dispersed ink spontaneously gather itself into a compact drop. Glass shatters, heat flows from hot to cold, and gases diffuse, all in a preferred temporal direction. This directionality is captured in thermodynamics by the second law, which states that in a closed system entropy tends to increase. While the underlying microscopic equations of motion are, in principle, time-reversal symmetric, the macroscopic world is structured around processes whose probability overwhelmingly favors states of higher entropy. This asymmetry is not an arbitrary feature of nature but a statistical regularity that constrains what kinds of patterns can persist, what kinds of systems can form, and how those systems can store and process information over time.
Entropy, in its thermodynamic sense, quantifies the number of microstates compatible with a macrostate. Low-entropy configurations correspond to highly ordered arrangements that occupy a very small region of the space of possibilities, whereas high-entropy configurations occupy vastly more. The arrow of time emerges because it is overwhelmingly more probable for a system to evolve from a special, finely tuned, low-entropy state toward more generic, high-entropy ones than the other way around. When we observe an ordered structure persisting over time—whether a living organism, a crystal lattice, or a stable planetary orbit—we are, in effect, seeing a local deviation from the typical entropic drift of the universe as a whole, usually sustained by flows of energy and matter that export entropy to the environment. In this way, even the capacity of a system to maintain its internal organization is tightly linked to the thermodynamic conditions that define its temporal environment.
This physical understanding of time asymmetry can be extended to cognitive systems by viewing brains as special kinds of thermodynamic devices. Brains are not closed systems; they are open, dissipative structures that maintain their low-entropy organization by consuming energy and exporting entropy to their surroundings. Neural activity, synaptic plasticity, and large-scale connectivity patterns all depend on metabolic resources, ion gradients, and molecular turnover. The very fact that a brain can sustain coherent dynamics over seconds, minutes, or years, instead of disintegrating into thermal noise, is due to continuous energetic support. In this respect, the arrow of time at the neural level is inseparable from the broader thermodynamic arrow: cognitive processes unfold in a temporally directed manner because the physical substrate that implements them is embedded in, and constrained by, a world whose macroscopic evolution favors increasing entropy.
Cognitive systems, however, do more than passively exist within this thermodynamic flow; they exploit temporal structure. Organisms survive by anticipating what will happen next and acting accordingly, and this anticipatory capacity depends on regularities in how causes and effects are ordered in time. Cause precedes effect, signals travel at finite speeds, and interventions now influence states later, not earlier. These asymmetries define a causal arrow that, in practice, aligns with the thermodynamic arrow: low-entropy configurations in the past constrain what is possible in the present, and present states in turn shape higher-entropy futures. Cognitive architectures, especially those capable of learning and memory, are tuned to this structure. They internalize probabilistic dependencies that reflect how the environment evolves and how actions lead to consequences in a temporally ordered way.
From this perspective, information processing in the brain can be seen as a specialized form of entropy management. When an organism encodes a regularity in its environment, it effectively compresses a wide range of possible sensory patterns into a more compact, structured internal representation. This compression reduces uncertainty about future inputs, allowing the system to exploit predictable features and ignore irrelevant noise. Yet this reduction in informational entropy at the cognitive level is possible only because of ongoing thermodynamic work: synapses are strengthened or weakened via energy-consuming biochemical cascades, neurons maintain concentration gradients through ATP-dependent pumps, and glial cells support the continuous turnover of cellular components. Cognitive order is purchased by physical disorder elsewhere, a local decrease in one form of entropy enabled by an increase in another.
The relation between physical and cognitive arrows of time becomes even clearer when considering learning. Learning mechanisms adapt internal parameters so that future sensory inputs are better anticipated. This adaptation presupposes a directed temporal flow from past data to future behavior. Data acquired at one time are used to adjust internal structures that will shape expectations and actions later. The inverse scenario—where future data determine present structures without any forward propagation—is neither biologically plausible nor consistent with observed thermodynamic constraints, absent exotic assumptions about retrocausality. Instead, the standard picture is of an organism incrementally aligning its internal dynamics with patterns extracted from past interactions, thus embedding the directionality of experience into its own architecture.
On longer timescales, evolution itself exemplifies a thermodynamic and informational arrow intertwined. Populations of organisms, through differential survival and reproduction, accumulate adaptations that reflect regularities in their environments. These adaptations are encoded in genetic and epigenetic structures, which bias the developmental pathways available to future individuals. In this way, information about historical environmental conditions is inscribed in the material of the organism, narrowing the range of viable phenotypes and behaviors. The process is again unidirectional: genetic changes propagate forward, and the space of forms explored is constrained by past selection. The thermodynamic costs of replication, repair, and reproduction are substantial, but they enable the persistent encoding of structure that carries the signature of a temporally ordered history.
When brains are viewed as products of such evolutionary and developmental processes, their capacities for perception, memory, and decision-making emerge as refined tools for harnessing temporal asymmetry. Neural circuits have been shaped to exploit the fact that the world exhibits lawful regularities in how states follow one another, and that interventions can alter future outcomes. Every act of inference or control takes for granted that the present is a hinge between a relatively fixed, recorded past and a range of more open, modifiable futures. This conceptual hinge aligns with the physical arrow of time defined by thermodynamics: information from the low-entropy past is stored and re-expressed in ways that allow the system to navigate toward specific regions of its high-entropy future, rather than drifting aimlessly.
At a finer grain, the temporal structure of neural dynamics itself reflects the underlying arrow of time. Spiking patterns, oscillations, and synchronization events are temporally ordered sequences that presuppose an orientation from earlier to later states. Many neural mechanisms, such as spike-timing-dependent plasticity, are explicitly time-asymmetric: synaptic strengths change differently depending on whether pre-synaptic spikes precede or follow post-synaptic spikes. This temporal asymmetry is not merely a detail of implementation but a core feature that allows neural circuits to encode causal relations and temporal contingencies. The rules governing how synapses change assume that earlier activity can be a cause of later activity, and they exploit this flow to embed directional relationships in the connectivity of the network.
Information-theoretic analyses of neural and behavioral data often reveal a similar bias toward the future. Measures such as directed information, transfer entropy, and Granger causality are designed precisely to capture how information flows from past states of one process to future states of another. When applied to neural populations or brain–body–environment interactions, these tools typically identify a predominant direction of influence that tracks the physical arrow of time: prior neural activity helps explain subsequent sensory inputs, motor outputs, or internal states more than the reverse. This directionality is not an arbitrary artifact of modeling; it expresses the deep alignment between the thermodynamic constraints on physical processes and the functional organization of cognitive systems that operate within those constraints.
By situating perception, action, and learning within this broader context, cognitive systems can be understood as entities whose very existence and operations presuppose a particular temporal structure. Their physical substrates rely on energy dissipation and entropy production, while their informational architectures capitalize on the consistent ordering of causes and effects. The thermodynamic arrow of time supplies the backdrop against which cognitive arrows—those of learning, inference, and control—can emerge and stabilize. Without the underlying physical asymmetry, the notions of past and future that guide cognitive processes would lose their operational meaning, and the capacity of minds to orient themselves in time would be fundamentally undermined.
Entropy, information, and the direction of inference
To understand how inference becomes temporally directed, it is useful to distinguish between thermodynamic entropy and informational entropy while recognizing that they are not unrelated. Thermodynamic entropy, rooted in statistical mechanics, counts the number of microscopic configurations compatible with a macroscopic state. Informational entropy, as defined by Shannon, quantifies uncertainty over possible messages or states. Both notions invoke a measure of multiplicity: the more ways something can be, the higher its entropy. Yet they apply to different descriptive levels—one to physical microstates, the other to probability distributions used by an observer or system. Cognitive systems bridge these levels: their physical implementation is constrained by thermodynamics, while their functional role is to manage informational entropy about the world they inhabit.
Whenever a cognitive system forms beliefs about its environment, it is implicitly working with probability distributions and thus with informational entropy. Before observing a stimulus, the system entertains a range of possibilities; the wider and more uniform this range, the higher its uncertainty. Sensory input, by ruling out some possibilities and favoring others, reduces this uncertainty in a context-dependent way. From an information-theoretic perspective, perception is the process of updating a prior distribution over hidden causes into a posterior distribution conditioned on new data. The direction of this update—from prior to posterior, from before to after observation—builds a temporal orientation into inference. The system never starts from the posterior and moves to the prior; the flow is consistently aligned with the ordering of sensory events in time.
This alignment is captured elegantly in the picture of the bayesian brain. On this view, neural systems implement approximate Bayesian inference, maintaining structured priors about the causes of sensory inputs and continuously revising them as evidence accumulates. Priors embody knowledge distilled from past interactions—developmental, experiential, and evolutionary—about which states of the world are more likely than others. When new data arrive, the system combines these priors with the likelihood of the data under different hypotheses to obtain a posterior belief state. This process defines a clear direction of inference: older information is encoded in priors, newer information enters as likelihoods, and the posterior becomes the new basis for subsequent predictions. The arrow of time is thus mirrored in the computational flow from prior to posterior, not only in physics but also in the internal logic of probabilistic reasoning.
Information-theoretic entropy plays a central role in this process by quantifying how “spread out” the system’s beliefs are at each stage. High-entropy priors represent broad uncertainty, where many hypotheses are considered plausible. Observations typically narrow this space, lowering entropy by concentrating probability mass on a smaller set of hypotheses that better explain the data. However, this reduction in informational entropy is not free: it requires physical work to acquire, encode, and maintain the structures that support refined priors and efficient updates. Synaptic plasticity, circuit reconfiguration, and long-term memory storage all instantiate reductions in uncertainty about the environment, but they do so through thermodynamically costly operations that consume energy and produce heat. This coupling ensures that the direction of inference—from uncertain to more informed—remains tethered to the thermodynamic arrow that underwrites all physical change.
The direction of inference can also be framed in terms of constraints on possible probability updates. In a temporally ordered world, information about earlier states is typically more accessible than information about future states. Signals travel forward, records accumulate, and traces of past events are etched into physical media. Cognitive systems exploit this asymmetry by building models whose structure assumes that data come from the past and bear on the future only via intervening causal processes. When a system performs a prediction, it is effectively propagating its current probabilistic beliefs forward along the expected causal structure, not backward. While formal Bayesian calculus allows for inferences about both past and future given current data, the operational sequence in real organisms privileges forward-looking use: updated posteriors guide subsequent action and expectation, anchoring the experiential sense of moving from a known or recorded past toward an uncertain future.
Information-theoretic tools help to make this directionality explicit. Measures such as mutual information are time-symmetric, treating dependencies between variables without regard to their temporal order. In contrast, quantities like transfer entropy and directed information are defined to capture asymmetries in how past states of one process help predict future states of another. When applied to brain activity, these measures typically reveal a predominant flow of information from earlier neural states to later ones, and from sensory and contextual inputs toward motor and decision-related areas. The fact that these directed measures are informative at all depends on the underlying temporal asymmetry of the world: environmental states leave signs that persist, and those signs can be used to shape later behavior. Without a background arrow of time, the very distinction between “incoming evidence” and “outgoing predictions” would blur.
The relationship between entropy and inference also clarifies why retrocausality is rarely invoked in cognitive explanations. In standard physical and statistical frameworks, probability distributions are updated as new data become available, not in anticipation of data that have not yet occurred. While one can formally condition on future observations when analyzing complete datasets, a real system operating online cannot access information that does not yet exist in its causal past. Its priors are constructed from thermodynamically realized processes—past events that have already unfolded and left stable traces. The idea of future measurements determining present internal states without any mediating forward process would require a breakdown of the usual temporal boundary conditions. As long as cognition is grounded in physically realized memories and signals, inference will reflect the asymmetries encoded in those physical substrates.
The way entropy shapes the direction of inference can also be seen in how systems allocate attention and resources. Environments vary in their predictability: some aspects are stable and low-entropy, others are volatile and high-entropy. Cognitive architectures that aim to minimize long-term uncertainty must decide which regions of state space are worth modeling in detail and which can be treated as noise. This selection process is inherently forward-oriented. A system does not merely recount past regularities; it prioritizes those that will reduce future prediction errors most effectively. Thus, informational entropy is not only a measure of current uncertainty but also a guide for where to invest computational and energetic effort. The temporal asymmetry of action—intervening now to influence what happens later—channels these investments toward patterns that are expected to persist into the future rather than toward fleeting coincidences that will soon dissolve.
Within predictive processing frameworks, this interplay is captured by the iterative cycle of prediction and error correction. The system generates top-down predictions based on its current generative model, compares them to incoming sensory data, and computes prediction errors. These errors carry information about the mismatch between expected and observed inputs, effectively measuring residual uncertainty or surprise. Reducing these errors corresponds to lowering the informational entropy of the system’s beliefs about the causes of its sensations. Notably, the cycle itself is temporally skewed: predictions are cast forward, errors arise only after new data arrive, and adjustments to the model prepare it for subsequent rounds of prediction. Each step presupposes an ordered sequence in which causes give rise to observable effects that can be evaluated only after they occur.
At a broader scale, the accumulation of knowledge can be seen as a cascade of entropy reductions in internal models, synchronized with entropy increases in the environment due to energy use and waste production. Each refinement of a cognitive system’s priors—whether through individual learning or across generations via evolution—narrows the range of hypotheses it must entertain to deal effectively with typical situations. These narrowed priors, in turn, enable more efficient use of sensory data, since fewer bits are needed to discriminate among a smaller set of plausible causes. The net result is a progressive alignment between the structure of the system’s internal probability distributions and the stable regularities of its niche. This alignment deepens over time but never reaches total certainty, because both the world and the system remain subject to fluctuations and ongoing thermodynamic drift.
Thus, entropy, information, and the direction of inference are tightly interwoven. The thermodynamics of physical substrates ensures that traces of the past can be stably stored while future states remain inherently more open. Informational entropy quantifies the uncertainty that cognitive systems face as they navigate this openness. Bayesian updating, predictive processing, and related inferential schemes give this navigation a definite orientation, from priors grounded in earlier events toward posteriors that guide future-oriented predictions and actions. The arrow of time is therefore not only a feature of physical processes but is also inscribed in the very logic by which cognitive systems reduce uncertainty and make sense of the worlds they inhabit.
Predictive processing and temporal asymmetry
Predictive processing offers a systematic way to connect the arrow of time in thermodynamics with the temporal structure of cognition. On this view, the brain is a hierarchically organized system that constantly generates predictions about incoming sensory data, compares those predictions to the actual input, and uses the resulting discrepancies—prediction errors—to update its internal model. The flow of computation is temporally ordered: priors are used to predict what will happen next, sensory evidence arrives afterward, and only then can the brain evaluate and revise its model. This sequence cannot be inverted without invoking exotic notions like retrocausality, because it relies on the simple fact that causes precede their observable effects and that information about the environment becomes available only once physical processes have unfolded.
In a predictive processing architecture, higher levels of the neural hierarchy encode more abstract, slowly changing causes of sensory input, while lower levels encode more transient, rapidly fluctuating details. Information flows in two main directions. Top-down signals convey prior expectations and generative hypotheses about what the world is currently doing; bottom-up signals convey prediction errors that register mismatches between those expectations and actual sensory states. Crucially, both flows are oriented in time. Priors are projected forward—from an already-constructed model into hypothetical near-future sensory states—while prediction errors arrive later, once those hypothetical states are confronted with real data. This is not a neutral, symmetric exchange of information; it is a temporally skewed loop that embodies a cognitive arrow of time aligned with physical causation.
The temporal asymmetry becomes clearer when we consider how the generative model is learned and maintained. The parameters that encode beliefs about hidden causes are not set instantaneously; they are gradually tuned through synaptic plasticity and structural reorganization driven by accumulated prediction errors. Each synaptic change is anchored in a history of discrepancies between predictions and outcomes. The brain’s current priors thus embody a sedimented record of past interactions—an integrated summary of prior prediction errors that have been thermodynamically realized as changes in the physical substrate of neural tissue. This embedding of history in structure is one way in which thermodynamics and information processing intersect: entropy produced in learning processes leaves behind low-entropy organizational patterns that bias future predictions.
Temporal asymmetry is also reflected in how prediction errors are used. Once a given round of prediction and comparison has occurred, the resulting error signals are not retroactively applied to earlier states of the system. They propagate forward in time, altering subsequent neural dynamics and shaping the next cycle of predictions. The brain cannot go back and change the priors that it had at a previous moment; it can only adjust the priors it will deploy in the future. This forward-only revision is analogous to the way macroscopic physical systems evolve toward higher entropy without spontaneously retracing their previous microstates. In both cases, the evolution of the system is constrained by an irreversible accumulation of changes that have already occurred and cannot simply be undone.
Within this framework, the temporal direction of inference is inseparable from the way sensory signals are generated and received. Sensory organs are physically coupled to external processes that unfold over time: photons bounce off objects before they reach the retina, pressure waves propagate before they trigger cochlear transduction, and mechanical forces impinge on the skin before they activate somatosensory receptors. The brain only ever receives the “outcome” of these causal chains. Predictive processing capitalizes on this structure by assuming that incoming data are consequences of hidden causes and by using generative models to infer those causes. The very notion of “hidden cause” presupposes that there is something that happened in the world that can explain the sensory data now arriving. This explanatory direction—from earlier cause to later evidence—builds temporal asymmetry into the representational scheme itself.
Another way to see the role of temporal asymmetry is through the treatment of uncertainty and entropy in predictive models. At any given moment, the brain’s generative model specifies a probability distribution over possible sensory inputs, conditioned on its current beliefs about hidden states. The entropy of this distribution measures how uncertain the system is about what will happen next. When new data are observed, prediction errors indicate how much of that uncertainty has been resolved or how much residual unpredictability remains. The system can then adjust the precision it assigns to different streams of input, effectively learning which aspects of the environment are stable and which are volatile. This process of tracking and reducing uncertainty is inherently future-directed: the purpose of refining priors is to improve the accuracy of upcoming predictions and minimize future surprise, not to re-describe the past for its own sake.
Temporal asymmetry is also implemented at the level of neural mechanisms that encode the relative timing of events. Many plasticity rules are explicitly time-sensitive, such as spike-timing-dependent plasticity, where synapses are strengthened if a presynaptic neuron fires just before a postsynaptic neuron and weakened if the order is reversed. These rules effectively distinguish potential causes from potential effects based on their temporal ordering. In a predictive processing system, such mechanisms support the learning of directed relations between events: patterns that reliably precede others become treated as predictors, and the brain’s generative model comes to expect that certain configurations will be followed by specific consequences. This learned directionality is not inferred from static co-occurrence alone; it depends on the fine-grained temporal structure of neural activity that tracks the unfolding of events in time.
The hierarchical structure of predictive processing further amplifies temporal asymmetry by operating across multiple timescales. Higher cortical areas often encode slow-changing contextual factors—such as environmental regularities, social norms, or bodily needs—that persist over seconds, minutes, or longer. Lower areas encode rapid fluctuations, such as momentary sensory details and fine motor commands. Predictions from higher levels thus span more extended temporal windows, constraining what is likely to occur not just in the next instant but across an entire sequence of events. Prediction errors at lower levels are evaluated against this broader temporal backdrop and are integrated over time to refine higher-level priors. The result is a nested set of arrows of time, with each level maintaining its own characteristic temporal horizon while remaining anchored to the same underlying thermodynamic flow.
Active inference, a natural extension of predictive processing, makes temporal asymmetry even more explicit by treating action as another route to minimizing prediction error. Instead of merely adjusting internal models to fit incoming data, an agent can also act on the world so that the data better fit its predictions. Actions are selected to fulfill prior expectations about future sensory states: the system anticipates what it “expects” to sense and then generates movements that bring about those expected sensations. This requires a robust sense of temporal ordering. Motor commands issued now are intended to influence sensory inputs that will occur later, and evaluations of whether those commands were successful depend on comparing expected and actual outcomes after the fact. The loop of intending, acting, and evaluating is thus inherently future-oriented, tied to a forward-moving sequence of cause and effect.
Computationally, the bayesian brain formulation underwrites this whole framework by casting perception and action as approximate Bayesian inference under a generative model. Priors encode assumptions about how hidden states evolve over time, often via dynamical models that specify transition probabilities from one moment to the next. Likelihoods encode assumptions about how those states give rise to observations, again in a temporally ordered manner. Temporal asymmetry is embedded in these generative assumptions: state transitions are defined from earlier to later time points, and the emission of sensory signals is modeled as a consequence of current states, not future ones. While Bayesian calculus allows for smoothing and backward inference over complete time series, the online algorithms that real brains must implement necessarily operate in a stepwise, forward fashion, constrained by when information actually becomes available.
Temporal asymmetry also shapes how predictive systems handle volatility and change. Environments are not perfectly stationary; the statistical structure governing events can drift over time. Predictive processing models often incorporate mechanisms that track the rate of change in underlying causes, adjusting learning rates or precision estimates accordingly. When a system detects that its predictions are persistently inaccurate, it can infer that the environment has changed and respond by updating its priors more rapidly. This capacity to detect shifts and reconfigure expectations depends on an ongoing comparison between long-term regularities and short-term deviations, a comparison that unfolds only as new data accumulate. The system cannot anticipate such changes before they manifest; it infers them retrospectively and then uses that inference to alter its future predictive stance.
Empirically, temporal asymmetry in predictive processing is evident in phenomena such as sensory adaptation, mismatch negativity, and anticipatory neural activity. In sensory adaptation, neurons reduce their responses to repetitive, predictable stimuli, effectively discounting information that carries little new surprise. This shows that current processing is shaped by a history of exposure and that the system expects stability unless contradicted. In mismatch negativity paradigms, the brain responds more strongly to unexpected deviations in a regular sequence, signaling prediction errors when temporal patterns are violated. Anticipatory signals in motor and premotor areas appear before movements and expected sensory outcomes, illustrating the forward projection of predictions into the near future. All of these empirical signatures presuppose a unidirectional flow of time, where what has already occurred shapes how the system prepares for what is about to occur.
The alignment between predictive processing and the thermodynamic arrow of time underscores that cognitive temporal asymmetry is not merely a matter of subjective experience or high-level narrative. It is rooted in the physical implementation of inference and control. The energy required to maintain gradients, propagate signals, and modify synapses is irreversibly dissipated; each act of prediction, comparison, and learning leaves a thermodynamic trace. Just as entropy production in physical systems favors a progression from special, low-entropy states to more generic, high-entropy ones, predictive brains move from relatively unstructured priors toward increasingly specialized models that reflect the regularities of their niche. This dual alignment—between physical and inferential arrows of time—supports a coherent picture in which cognitive systems are temporally embedded engines of prediction, exploiting and depending upon the very same asymmetries that structure the physical world.
Memory, causation, and the experience of time
Memory is the faculty through which cognitive systems convert the physical irreversibility of events into a structured sense of before and after. Every episode an organism undergoes leaves traces in its nervous system, from transient changes in firing patterns to long-lasting modifications of synaptic connectivity and gene expression. These traces are not neutral records; they are oriented in time. They encode information about what has already happened and are deployed to constrain what is likely to happen next. In this way, memory transforms the thermodynamic arrow of time into a cognitive arrow: the accumulation of physical changes that cannot be undone becomes the basis for a directed flow of information from past to present to future.
The physical substrate of memory makes this temporal orientation especially clear. Synaptic plasticity, structural remodeling, and molecular consolidation all depend on metabolic processes that dissipate energy and increase entropy elsewhere in the system and its environment. Once a synapse has been potentiated or depressed in response to activity, the biochemical cascades that produced that change cannot simply be “run backward” to restore an earlier state without additional, carefully orchestrated work. Even when forgetting occurs, it follows its own irreversible trajectory, driven by degradation, noise, and new learning rather than a precise reversal of earlier modifications. Memory thus rides on the same thermodynamic asymmetry that governs other macroscopic processes: it is easier to accumulate and transform traces than to erase them in a way that recreates a previous microstate exactly.
From an informational perspective, memory can be understood as the storage of compressed regularities extracted from experience. The brain does not attempt to keep a perfect, bit-by-bit replica of the past; instead, it distills patterns that are likely to matter for future prediction and control. This compression reduces informational entropy over relevant variables: many distinct episodes get summarized into a smaller set of generalized expectations and schemas. Yet that reduction in informational entropy requires thermodynamic work. Forming long-term memories engages processes such as protein synthesis, synaptic growth, and reorganization of network connectivity, all of which consume energy and generate heat. Memory therefore exemplifies a trade-off: cognitive systems pay a physical cost to create low-entropy, structured internal states that will later guide behavior.
In the bayesian brain framework, memory is naturally cast in terms of priors. What an organism has experienced—and, on even longer timescales, what its ancestors have undergone—is crystallized into prior probability distributions over hidden causes and likely outcomes. These priors express a temporally accumulated bias: they encode what has tended to be true, given a history of interactions with a particular environment. When new data arrive, they are evaluated against these priors, and the resulting posteriors feed forward into subsequent rounds of expectation and action. The direction of this process mirrors the direction of memory: past events shape priors, priors constrain present perception, and present perception updates the priors that will influence future processing. No step in this chain requires, or admits, genuine retrocausality; it is always the case that earlier states inform later ones through stored traces.
The connection between memory and causation emerges most sharply when considering how organisms infer what caused a given outcome. The environment continually presents ambiguous sensory signals that could have arisen from multiple possible sources. To disambiguate them, cognitive systems rely on stored knowledge about which configurations tend to precede others and how interventions change what happens later. Causal learning mechanisms track asymmetric dependencies: if one event reliably comes before another and manipulations of the first systematically alter the second, the system infers a directed relation. Memory is essential here because detecting and exploiting such asymmetries requires integrating information across time. Without a record of how patterns have unfolded, a system would see only isolated snapshots, unable to distinguish cause from mere correlation.
At the neural level, mechanisms like spike-timing-dependent plasticity provide a bridge between temporal order and causal attribution. Synapses are strengthened when presynaptic activity reliably precedes postsynaptic firing, and weakened when the order is reversed or inconsistent. This time-sensitive rule effectively encodes a primitive notion of “if A then B”: neurons whose activity anticipates that of others become better positioned to drive them in the future. Over many repetitions, such plasticity sculpts circuits that embody directed associations aligned with the experienced order of events. The causal arrow learned by the network is therefore grounded in the temporal arrow imposed by the environment and implemented through irreversible changes in synaptic strengths.
Memory also underwrites the distinction between events that merely co-occur and events that can be controlled. An organism’s sense of agency depends on its ability to differentiate outcomes that follow from its own actions from those that would have occurred regardless. Achieving this distinction requires comparing what happens when it acts with what happens when it does not, across multiple instances. These comparisons, in turn, require retaining information about prior episodes: which motor commands were issued, which sensory consequences ensued, and how tightly the two covaried. As patterns of reliable contingency accumulate, the organism learns causal models in which its own states occupy a privileged position as potential initiators of change. The subjective feeling of “I did that” rests on this memory-based capacity to track asymmetric dependencies between internal signals and external events.
The experience of time itself is deeply shaped by how memories are formed, organized, and accessed. Episodic memory, in particular, presents past events as belonging to a personal timeline with a definite order: one can recall that breakfast came before the morning meeting and after waking up, rather than as a jumble of disconnected scenes. This temporal ordering is supported by neurobiological systems, such as the hippocampus and associated cortical networks, that encode both content and context, including temporal context. Sequences of neural activity—sometimes replayed during rest or sleep—provide a scaffold for reconstructing the serial structure of experience. The fact that such sequences can be replayed forward or, in some cases, backward does not imply a reversal of the arrow of time; it reflects a flexible retrieval process operating over traces that were originally laid down in one direction.
Moreover, different kinds of memory impose different temporal horizons on experience. Working memory maintains information over short intervals, allowing an organism to integrate fleeting sensory inputs and plan immediate actions. Semantic memory accumulates more stable knowledge about the world’s enduring regularities. Procedural memory encodes skills that unfold as well-practiced sequences over characteristic timescales. Together, these memory systems define how far into the past an organism can reach and how far into the future it can project. A richer and more reliable memory allows longer chains of causal inference and more extended forms of planning, effectively stretching the lived temporal window within which the organism can situate itself.
The organization of memory also shapes how time is subjectively partitioned. Events that share thematic or causal coherence tend to be encoded as unified episodes, while shifts in context, goals, or outcomes mark boundaries between episodes. These boundaries influence later judgments about temporal distance and duration: periods filled with many distinct, memorable events may later seem longer than periods that were, in real time, equally extended but more homogeneous. Here again, the experience of time reflects properties of memory storage and retrieval rather than a direct readout of an external temporal metric. Yet the asymmetry remains: memory points backward, and the sense of an unfolding life narrative depends on being able to place remembered events in a coherent ordered structure that stretches from an earlier self toward a not-yet-realized future.
The coupling between memory and causation also constrains how far cognitive systems can, in practice, reason about time in non-standard ways. While formal models may allow computations that condition on “future data,” biological systems are limited to information that has already entered their causal past and been encoded in some substrate. Prospective simulations—imagining what might happen under different scenarios—are built by recombining and extrapolating from existing memories, not by accessing future facts. When an organism appears to “remember the future,” as in strong anticipatory behavior, it is actually drawing on previously learned regularities that make certain forthcoming events highly predictable. The brain’s prediction machinery can thus generate vivid experiences of possible futures, but these experiences are anchored in stored causal patterns, not in any actual reversal of the arrow of time.
Viewed through the lens of thermodynamics, memory can be seen as a local counterforce to the global trend toward disorder. By expending energy, cognitive systems create and maintain ordered, low-entropy records that track the structure of their environment and their own interactions with it. These records allow them to resist some of the immediate consequences of environmental unpredictability, navigating instead by learned regularities and causal expectations. Yet this local ordering is always provisional and limited. Memories decay, are overwritten, or become mismatched to changing circumstances, and maintaining them requires continuous work. The direction of forgetting, like that of learning, is not arbitrary; it reflects ongoing adjustments to a world whose macroscopic evolution relentlessly favors higher entropy.
In everyday cognition, the interplay between memory and causation manifests in how agents construct explanations and narratives. When something unexpected happens, people search their memories for prior events that could plausibly have led to it, arranging these in a chain of reasons that run from earlier to later. Even when such explanations are post hoc and partial, they rely on a store of temporally indexed information and on intuitive models of how causes operate over time. These models are often biased—for instance, overemphasizing recent events or salient interventions—but their basic structure presupposes a unidirectional mapping from past conditions to present outcomes. The felt coherence of a story about “what happened” depends on this mapping and would collapse if events could be freely rearranged without regard to their order.
Ultimately, memory, causation, and the experience of time are intertwined aspects of a single adaptive strategy. By storing traces of past interactions in a physically irreversible manner, cognitive systems create a reservoir of information that can be used to infer causal structure and guide future behavior. The thermodynamic arrow of time ensures that such traces accumulate in one direction only, while the inferential machinery of the brain translates that accumulation into priors, predictions, and judgments about what leads to what. The subjective sense of moving through time—from a remembered past, through a lived present, toward an anticipated future—emerges from this ongoing interaction between irreversible physical processes and the organized patterns of memory they leave behind.
Implications for cognition, agency, and free will
If cognition is grounded in predictive processing within a universe governed by thermodynamic asymmetries, then agency and free will must be understood in ways that respect both the arrow of time and the constraints of physical implementation. Brains are not detached observers that merely register what happens; they are embodied controllers that use predictions to sculpt the flow of events within the narrow corridor allowed by physics. Their generative models contain priors over possible futures, and actions are selected to make certain predicted trajectories more likely than others. Agency, on this view, is not the ability to act without causes but the capacity of a system to exploit causal structure: to use its internal model to bias which among many thermodynamically permissible futures actually comes to pass.
This perspective reframes freedom in terms of model-based control rather than metaphysical exemption from laws of nature. A system enjoys a form of freedom to the extent that it can evaluate alternative courses of action, assess their likely consequences, and choose among them in a way that is sensitive to its preferences and constraints. All of these capacities depend on the same asymmetries that underwrite the arrow of time. Evaluating alternatives presupposes memory of past outcomes and the ability to simulate their downstream effects; assessing consequences presupposes causal knowledge learned from temporally ordered experience; choosing among options presupposes mechanisms for comparing predicted futures and stabilizing one course of action long enough to execute it. None of this involves retrocausality. Rather, it involves forward-looking deployment of information that has been accumulated irreversibly from the past.
In the bayesian brain framework, these capacities can be expressed in terms of probability distributions over action policies and their expected outcomes. Priors encode the organism’s default expectations and preferences—for instance, to maintain homeostasis, avoid injury, and seek resources. When new sensory evidence arrives, it updates not only beliefs about the current state of the world but also beliefs about which actions would minimize expected prediction error or, in free energy formulations, expected surprise over extended time horizons. Agency, in this sense, is the expression of a constantly re-optimized policy under uncertainty: a sequence of decisions that emerges from the interplay between prior constraints, incoming data, and inferred consequences. Freedom is not absence of constraints but the richness and flexibility of the constraint landscape within which such probabilistic control is exercised.
Thermodynamics enters this story in two ways. First, it sets the energetic budget within which cognitive control must operate. Any process that stabilizes low-entropy internal order, such as maintaining reliable memories or implementing complex planning, incurs a metabolic cost. Brains cannot entertain arbitrarily many alternatives or simulate indefinitely long futures; they must balance the benefits of improved prediction against the costs of additional computation. Second, thermodynamics sets the reliability of physical implementations of decisions. Motor commands must be transmitted through noisy, dissipative channels in muscles and the environment. The more finely an agent tries to control its surroundings, the more sensitive its strategies become to fluctuations and degradation. Agency is therefore always bounded: it arises from and is limited by the trade-offs between informational complexity, energetic expenditure, and robustness to noise.
These constraints do not eliminate meaningful notions of responsibility or choice; they shape them. Because agents are predictive systems, their actions depend not only on immediate stimuli but also on long-standing priors shaped by learning and development. When we say that someone “could have acted differently,” in this framework we are implicitly referring to the possibility that, under slightly different priors or under a modified evidential context, the same physical organism would have selected a different action policy. Interventions like education, therapy, or training can be understood as attempts to alter those priors and the inferential machinery that deploys them, thereby changing the distribution of likely future actions. Responsibility then tracks the degree to which a person’s internal models, formed over time, systematically mediate between situational inputs and behavioral outputs.
This mediation is temporally extended and deeply tied to memory. Choices at any given moment reflect a layered history: genetic dispositions, early developmental experiences, habits, and recently acquired information all contribute to the priors that shape present inference. From a thermodynamic standpoint, this history is written into the structure of the nervous system as a set of low-entropy configurations stabilized at continuing energetic cost. From a cognitive standpoint, it appears as stable traits, values, and tendencies. Agency is therefore not a property of an instant but of a process: a temporally thick pattern in which an organism continually reuses its past to constrain its future. Holding an agent responsible is, in part, recognizing that this pattern is robust enough that future behavior can be influenced by adjusting the informational and social conditions under which it evolves.
At the same time, the arrow of time highlights why absolute, counterfactual freedom—freedom unconstrained by prior causes—is incoherent within a predictive framework. An agent cannot choose its own priors ex nihilo; those priors are the cumulative product of genetic variation, developmental contingencies, and prior actions. Nor can an agent, at a given moment, step outside the unfolding causal chain that determines which prediction errors it encounters and which policy it ultimately selects. What it can do, however, is model aspects of this chain, anticipate how different choices are likely to alter later states of itself and its environment, and act in light of that anticipation. Reflexive self-modeling—treating one’s own future states as objects of prediction—enables a more sophisticated, temporally extended form of control that approximates what is ordinarily meant by deliberative free will.
Deliberation, in this light, is a specialized mode of predictive processing in which internal simulations are given greater weight than immediate sensory inputs. Rather than passively allowing external events to drive prediction errors, the agent explores hypothetical scenarios, comparing their expected outcomes under different action policies. This mental time travel relies on memory, generative modeling, and the ability to sustain multiple competing trajectories in working memory long enough to evaluate them. The feeling of “choosing” among imagined futures corresponds to a process in which the system settles on a policy that best minimizes expected long-term entropy in domains it cares about—such as bodily integrity, social standing, or epistemic coherence—subject to its energetic and informational limits. Once a policy is selected, descending motor commands commit the organism to a particular branch of the future, closing off alternatives that had existed only as low-entropy internal possibilities.
The connection between entropy and agency also clarifies why some forms of behavior feel more autonomous than others. Actions that are heavily overdetermined by immediate stimuli, with little room for internal modeling or delay, involve minimal reduction in expected uncertainty: the environment effectively dictates the next state. By contrast, actions that emerge after prolonged reflection, weighing of options, and consultation of long-term goals reflect more extensive internal processing aimed at reshaping the range of future possibilities. These latter actions have a larger informational footprint: they engage broader swaths of the agent’s generative model, integrate more diverse memories, and often incur greater metabolic expense. The subjective sense of “owning” such decisions may track, in part, this increased involvement of internal, temporally deep structures in constraining the outcome.
This framework also offers a way to think about social and moral practices. Social environments are powerful shapers of predictive models: norms, laws, and institutions can be seen as external structures that alter the priors and likelihoods agents use when inferring the consequences of their actions. Punishments, rewards, and reputational feedback change the expected costs and benefits of different policies, effectively steering the space of predicted futures. When a society attributes responsibility, it is implicitly asserting that an individual’s generative model is sensitive to such shaping—that future behavior can be altered by changing incentives, information, or context. This assumption fits naturally with a view of agents as thermodynamically realized predictors whose behavior emerges from modifiable distributions rather than from fixed, inscrutable essences.
The alignment between cognitive and physical arrows of time underscores why notions of freedom that depend on causal influence from the future into the past are unnecessary. Predictive processing already explains how agents can be strongly oriented toward the future without invoking retrocausality. Anticipated outcomes exert their influence not by sending signals backward in time but by being represented now, as part of the agent’s generative model, and by guiding present choices. These representations themselves are products of past learning and ongoing thermodynamic processes in the brain. The sense that “the future is pulling us forward” is thus an experiential gloss on a forward-directed mechanism: priors, memories, and predictions interact in the present to shape actions that will, if all goes well, bring about states of the world that resemble the futures the agent currently prefers.
