Non-markovian minds and memory of the future

In classical cognitive science, mental processes are often modeled as Markovian: what the system does next is fully determined by its current state, independent of the detailed path that led there. This assumption simplifies modeling and computation, but it sits uneasily with the richness and persistence of human memory, habits, and context-sensitivity. Non-markovian cognition refers to mental processes whose future evolution depends not only on the present, but also on structured traces of the past that are neither compressible nor reducible to a single “current state.” In a non-markovian mind, the past is not merely summarized; it remains actively entangled with perception, thought, and action across extended timescales.

Temporal dependency in cognition is evident from the way meanings, expectations, and emotional responses accumulate over time. The interpretation of a sentence depends on words that appeared many clauses ago; the evaluation of a decision depends on a history of outcomes and regrets; the experience of a melody depends on an extended pattern of notes, not on any isolated tone. These are not just qualitative anecdotes. They reveal that what the mind does at any moment carries a residue of earlier processing that cannot be fully captured by a short-lived internal snapshot. Cognitive states are temporally thick: they inherit structure, constraints, and biases from sequences of prior states.

At the neurobiological level, temporal dependency is underwritten by overlapping timescales of neural dynamics. Ion channel kinetics, synaptic facilitation and depression, short-term and long-term plasticity, recurrent network loops, and neuromodulatory systems all store aspects of prior activity over milliseconds, seconds, minutes, hours, and years. Even if one could, in principle, write down a “complete” instantaneous neural state, that state would already embed a history of stimulation and plastic change. The present configuration of synaptic strengths, dendritic structures, and network connectivity is itself a physical record of past interactions. In this sense, the brain’s “now” is already saturated with “then,” making a purely Markovian description inadequate.

Working memory and long-term memory provide two salient examples of non-markovian structure. Working memory holds information across brief delays so that it can shape upcoming behavior: remembering a phone number, tracking the steps in a mental calculation, or keeping track of conversational context. This form of memory maintains specific patterns of neural activity or synaptic weights that directly constrain future processing. Long-term memory, in turn, provides a vast storehouse of facts, skills, and associations that were learned in the past but remain poised to influence recognition, decision, and imagination. Neither form of memory is just “background storage”; both are active resources that inject temporally extended information into ongoing cognition.

Temporal dependency also shows up in the way cognitive systems learn. Learning algorithms adjust internal parameters based on sequences of experiences, errors, and rewards. The resulting “priors” are not merely instantaneous beliefs but condensed histories of exposure to environmental regularities. When the brain encounters new input, it does not respond solely to what is present; it interprets that input through prior learned structure about what tends to co-occur, what tends to follow, and what tends to cause what. In this way, learning transforms temporal sequences into enduring constraints on how new information is processed, creating a deep non-markovian link between prior episodes and current inference.

A Markovian model can, in principle, be made more expressive by expanding its state space to include memory variables, but this move often misses the point. Non-markovian cognition is not just about adding a finite list of memory registers; it is about the fact that history is distributed across the hierarchical architecture of the nervous system and across the body and environment. Context from the past infiltrates perception and action through subtle changes in attention, readiness, and interpretation that cannot be localized to a single memory buffer. Temporal dependency is inherently multi-scale and often path-dependent: tiny variations in initial experiences can lead to divergent developmental trajectories, personalities, and cognitive styles.

Language processing offers a clear illustration of these temporal entanglements. Comprehension depends on syntactic and semantic expectations formed over the course of a sentence or discourse. When encountering a temporarily ambiguous phrase, the interpretation the mind settles on can hinge on words encountered seconds earlier, or even on background knowledge acquired years earlier. Eye-tracking and electrophysiological studies show that readers and listeners carry forward multiple potential parses and adjust them as new words appear, revealing a cognitive system that continuously maintains and revises structured dependencies over extended temporal windows. Such behavior is more naturally captured by non-markovian models that track trajectories of interpretation rather than single, self-contained states.

Similarly, decision-making is shaped by histories of reward, punishment, and feedback that extend well beyond the immediate task at hand. Risk attitudes, impulsivity, and perseverance are not fixed reaction rules applied to isolated choices; they depend on accumulated experiences of loss and gain, success and failure, trust and betrayal. Behavioral economics and reinforcement learning studies show that people’s current preferences can be heavily influenced by prior sequences of outcomes, framing effects, and reference points. These findings indicate that the mapping from present stimuli to choice is filtered through an evolving lattice of memory traces, undermining any attempt to characterize decisions solely in terms of current observable variables.

Emotion and affective dynamics further deepen temporal dependency. Emotional responses are not just reflexive outputs to present stimuli; they are modulated by mood, stress levels, circadian rhythms, and long-term dispositions shaped by developmental history, trauma, attachment patterns, and cultural learning. A remark that would be harmless in one context may provoke anger or sadness in another because of prior interactions, background tensions, or unspoken narratives shared over months and years. Emotional context acts as a slowly varying, history-laden parameter that tunes perception, appraisal, and action tendencies in ways that a Markovian snapshot cannot easily encode.

Motor control and skill acquisition likewise defy a strictly Markovian treatment. Skilled actions such as playing a musical instrument, typing, or athletic performance depend on motor programs honed over long periods of practice. These skills are embedded in sensorimotor loops where current kinematics depend on a history of micro-corrections and feedback. Anticipatory adjustments—such as preparing for a ball’s trajectory or synchronizing with a rhythm—draw on learned temporal regularities, not just immediate sensory input. In such cases, the system’s present configuration is the culmination of countless prior adjustments, and the unfolding of action trajectories depends on ingrained patterns that span extended time.

From a modeling perspective, non-markovian cognition can be handled with tools that explicitly incorporate temporal structure: recurrent neural networks with long memory traces, dynamical systems with structured phase space, and probabilistic models that track histories or latent processes evolving over time. These approaches move beyond one-step transition rules and instead represent cognition as the evolution of trajectories in high-dimensional spaces, where past positions continue to shape future directions. Such models acknowledge that human mental life is not a chain of independent, forgetful steps but a woven fabric of overlapping temporal threads.

Conceptually, recognizing the non-markovian character of cognition invites a shift from thinking of mental states as discrete, momentary snapshots to viewing them as segments of ongoing processes. What counts as “the current state” is already an integration of prior inputs, internal adjustments, and anticipatory setups for what is likely to come next. Temporal dependency is not a peripheral add-on to an otherwise memoryless core; it is a defining feature of how minds maintain coherence, identity, and adaptation across time.

Predictive processing and models of future memory

Predictive processing portrays the brain as a hierarchically organized inference machine that constantly attempts to minimize the mismatch between its predictions and incoming sensory signals. Rather than passively receiving data from the world, the system is thought to generate top-down expectations about the causes of sensory input, compare them to bottom-up evidence, and update its internal models in light of the resulting prediction errors. In this framework, perception is not a direct readout of the environment, but a controlled hallucination constrained by both sensory signals and prior beliefs about how the world typically behaves. This emphasis on ongoing inference naturally embeds temporal structure: predictions are not just about what is currently present, but about what is likely to happen next.

Crucially, predictive processing involves an interplay between fast-changing predictions at lower sensory levels and slowly evolving priors at higher, more abstract levels. Lower levels may predict immediate features such as edges, tones, or limb positions over milliseconds to seconds, while higher levels encode more stable regularities about objects, agents, and contexts that unfold over minutes, days, or years. These higher-level priors incorporate an extensive history of learning and experience, effectively condensing long temporal sequences into structured expectations. The result is a deeply non-markovian architecture: what is inferred at any moment depends not only on the present input but on a layered archive of past events that continues to shape the generative model.

In this view, memory and prediction are not separate faculties but two aspects of a single inferential process unfolding over time. The same generative model that explains current sensory input is also used to project forward into possible futures. Remembering a past event involves reactivating a pattern of latent causes that once generated particular sensory experiences; imagining the future involves recombining and extrapolating these causes into new configurations. Both operations draw on embedded temporal structure: sequences of causes, transition probabilities, and causal graphs that encode how states tend to follow one another. Memory is thus “future-oriented” not only in the sense that it guides future behavior, but in the stronger sense that what gets stored and consolidated is biased toward features that matter for future prediction and control.

Within a predictive processing scheme, so-called “future memory” can be understood as the brain’s capacity to maintain and refine detailed expectations about upcoming states, events, and contingencies that have not yet occurred. These expectations are not bare scalar forecasts, like a single predicted value of temperature tomorrow. They are rich, structured generative models specifying how sensory streams will likely unfold, how actions will modify those streams, and which contextual variables are relevant at which timescales. When we prepare for a conversation, a performance, or a potential conflict, we are in effect populating a space of possible futures and assigning them graded probabilities. These anticipatory states behave like memories of the future: they can be stored, refined, retrieved, and compared against what actually happens.

Neurobiologically, this anticipatory scaffolding is supported by neural dynamics that enable both short-lived and long-lasting traces of potential futures. Hippocampal “preplay” and “replay” phenomena in rodents illustrate this clearly: place cells that fire in sequences during navigation can also fire in compressed sequences ahead of movement, as if mentally simulating future trajectories. Similar mechanisms are observed during planning and decision-making, where neural activity transiently walks through alternative action paths and their likely outcomes. Rather than waiting passively for sensory input, these circuits preconfigure patterns of activity that embody predictions about future states of the world and the body, providing a concrete substrate for future memory.

Models of the “Bayesian brain” formalize these ideas by treating perception and cognition as approximate Bayesian inference in hierarchical generative models. Each level of the hierarchy maintains a probability distribution over latent causes, which is updated using Bayes’ rule in light of prediction errors from lower levels. Temporal dependencies are built into these models by specifying prior distributions over trajectories, not just static states. For example, dynamical priors might encode that objects tend to move smoothly, that sounds cluster into phonemes and words over characteristic timescales, or that social interactions follow recognizable scripts. In such models, the prior is already temporal: it encodes how conditions evolve, which implicitly requires “remembering” the path taken so far and projecting it forward.

From this perspective, a strictly Markovian representation becomes inadequate because the relevant priors often depend on extended histories rather than purely on the current state. Consider language comprehension. A Bayesian model that treats each word as conditionally independent given only the immediately preceding word will fail to capture long-range dependencies such as subject-verb agreement across clauses, thematic roles spanning an entire sentence, or discourse coherence across paragraphs. More realistic generative language models must maintain latent variables that persist and evolve over time, linking distant tokens and contexts. These latent trajectories make the effective state space history-dependent, pulling the system into a non-markovian regime even if the formalism superficially resembles a Markov chain augmented with hidden variables.

Predictive processing also reframes episodic memory as a constructive process tightly coupled to future-oriented inference. When recalling an episode, the system does not simply retrieve a static recording from storage; it re-infers a plausible pattern of causes that could have generated previous sensory inputs, guided by current goals and knowledge. This reconstruction is influenced by what the system now believes is typical, causal, or important. As a result, recollection is often biased toward features that are relevant for future prediction: causal relations, moral lessons, social rules, and risk-related cues may be amplified, while incidental details fade. Memory becomes an ongoing process of model refinement, where each act of remembering adjusts the generative model to better forecast similar situations in the future.

This view has direct implications for how we understand the malleability and distortions of memory. Classic phenomena such as hindsight bias, the “knew-it-all-along” effect, and the tendency to reshape past beliefs to align with current preferences can be seen as byproducts of predictive updating. When outcomes become known, the generative model is retrofitted so that the observed sequence appears more expected than it actually was. In doing so, the brain effectively overwrites earlier predictive states, creating a new narrative in which the eventual outcome seems inevitable. The psychological sense that “the future was already written” can thus arise from backward inference within a non-markovian generative model, not from any literal retrocausality in the physical world.

In everyday cognition, much of what we call planning, expectation, and concern is driven by continuously updated predictive states that function as standing “appointments” in the mind’s temporal landscape. Anticipating a meeting tomorrow, awaiting the results of a test, or rehearsing a difficult conversation all involve repeatedly reactivating and reshaping internal models of specific future situations. These models can acquire many of the phenomenological qualities of memory: rich imagery, emotional coloring, narrative structure, and a sense of familiarity. They can also influence present perception and action as strongly as, or more strongly than, memories of the actual past. For instance, anxiety disorders often revolve around intensely vivid negative future scenarios that dominate current attention and behavior, effectively acting as intrusive memories of things that have not yet happened.

In formal predictive coding implementations, the distinction between memory and prediction becomes even more blurred. The same error signals that adjust beliefs about current hidden causes also tune transition models that govern how those causes are expected to change over time. When the world consistently violates some aspect of our expectations, this is reflected not only in updated beliefs about present conditions but also in revised beliefs about temporal contingencies. Over repeated encounters, the system learns that certain sequences occur more often than others, that some cues reliably precede important events, and that specific actions lead to particular outcomes. These learned temporal relations are then implicitly “stored” in the generative model, so that future inputs are interpreted against a backdrop of deeply ingrained temporal structure.

Attention and precision weighting provide another route by which predictive processing supports non-markovian behavior. The system must decide how much weight to assign to current sensory input versus prior expectations, and this decision is itself shaped by a history of reliability, noise, and volatility. If recent experience has shown that a particular sensory channel is unreliable—say, vision in foggy weather—the system will down-weight prediction errors from that channel and lean more heavily on internal predictions. Conversely, in a novel or unstable environment, precision may be allocated to incoming data, prompting rapid updating. These precision estimates are not memoryless; they track patterns of stability and change over time, creating meta-level future memories about when to trust or distrust sensory evidence.

Generative models of action, such as active inference, extend predictive processing into motor control and decision-making. Here, the brain is seen as predicting not only sensory inputs but also proprioceptive consequences of its own actions. By minimizing prediction error, the system can bring about the predicted states, effectively using action to confirm its expectations. Planning becomes equivalent to selecting policies—sequences of actions—that minimize expected free energy, a quantity that balances predicted rewards, costs, uncertainty, and model mismatch over time. To evaluate such policies, the system must simulate their likely temporal unfolding, invoking a dense web of future-oriented inferences that depend on past learning. The selection of present actions thus rests on a tapestry of future memories: internally generated trajectories of possible worlds that are evaluated, compared, and pruned before any overt movement occurs.

In this light, non-markovian cognition is not a peripheral complication added onto an otherwise Markovian predictive engine; it is the natural outcome of a hierarchical bayesian brain equipped with deep temporal models. Each layer carries forward information not only about what is currently believed, but about how these beliefs are expected to evolve across time. Sensory data are interpreted through accumulated priors about temporal structure, and anticipated futures continuously modulate how the present is perceived and acted upon. What we remember, what we attend to, and what we expect are jointly determined by a single inferential architecture whose state at any moment encodes trajectories—past, present, and possible—rather than a thin, momentary snapshot of “now.”

Information theory and the limits of markovian minds

Information theory provides a rigorous vocabulary for analyzing how much of the past is carried into the present, and how much of the present constrains the future. Within this framework, a perfectly Markovian process is one where the future is conditionally independent of everything but the current state: once you know the state now, no additional knowledge of the past improves your prediction of what comes next. This property can be captured in terms of conditional mutual information: for a truly first-order Markov chain, the mutual information between the distant past and the immediate future, given the present, is zero. In other words, the present is a sufficient statistic for all temporally extended influence. Any residual information that the past has about the future, beyond what is encoded in the present, signals non-markovian structure.

When applied to cognition, these information-theoretic notions become powerful diagnostics of temporal dependency. The brain continuously transforms sensory inputs into internal states and actions, but those transformations can preserve, compress, or discard information about prior inputs. If mental processing were strictly Markovian at some level of description, then an appropriate “state” variable would fully screen off the future from the past. Yet measures like active information storage, transfer entropy, and excess entropy, applied to behavioral and neural time series, routinely reveal significant dependencies that span multiple time steps and even multiple scales. These quantities measure, respectively, how much the past of a process helps to predict its next state, how much influence one process exerts on another over time, and how much structure is embedded in an entire trajectory. Elevated values of these measures indicate that memory-like traces of prior states are still functionally relevant, contradicting a simple Markov assumption.

Excess entropy is particularly instructive. It quantifies how predictable a sequence is, over and above what can be captured by single-step transition probabilities. For a purely Markovian system of fixed order, excess entropy saturates once the model includes enough immediate history. But many cognitive and neural processes display long-range correlations that do not saturate quickly: prediction accuracy continues to improve as longer windows of past data are considered. This suggests that the effective state of the system cannot be encapsulated by a bounded set of recent variables. Instead, the system preserves structured information about its trajectory, distributed across synaptic weights, recurrent activity patterns, and slow neuromodulatory signals. Information theory thereby reveals a deep non-markovian character in neural dynamics, even when those dynamics appear noisy or irregular at short timescales.

The Markov assumption also plays a central role in common models of learning and decision-making. Standard reinforcement learning, for instance, often assumes a Markov decision process, where the probability of a reward and the next state depends only on the current state and action. Information-theoretic analysis shows that this assumption severely constrains what such agents can, in principle, learn and exploit. If important task features are temporally extended—such as sequences of cues that only gain meaning when assembled into a pattern—then a Markovian agent that neglects those extended histories leaves useful predictive information on the table. Quantitatively, the mutual information between distant past events and future outcomes, not captured by the current Markov state, sets an upper bound on how much performance can be improved by moving to richer, non-markovian representations.

From the perspective of the bayesian brain, Markov limitations can be understood in terms of priors over temporal structure. A Markov prior encodes that the most relevant information for prediction lies in the current state, whereas non-Markov priors permit extended dependencies among latent variables. Information theory clarifies the cost of adopting the simpler prior: if the environment exhibits long memory, then a Markov prior constitutes a systematic model mismatch. This mismatch manifests as persistent prediction errors that cannot be eliminated by any amount of parameter tuning within a Markovian family. These residual errors correspond to mutual information that the true process carries across long time spans, but which the model is structurally unable to represent. In this sense, the limits of markovian minds are not just practical but principled: certain patterns of temporal regularity are unlearnable under Markov constraints, no matter how much data is available.

Another way to see these limits is through rate–distortion theory, which studies how much information about an input process must be retained to achieve a desired level of performance on some task, such as accurate prediction or control. A cognitive system with finite resources cannot preserve all details of its sensory history; it must compress past information into manageable internal codes. A strict Markov model corresponds to an extreme compression strategy: discard everything but the present state. However, when the environment’s dynamics are non-Markovian, rate–distortion analysis shows that any compression that ignores longer histories incurs extra distortion in prediction. To achieve a given predictive accuracy, the system must allocate some of its limited channel capacity to carry forward selected aspects of the past. Information theory thus frames memory not as a luxury but as an optimal response to environmental complexity.

Computational mechanics, an information-theoretic framework for analyzing complex stochastic processes, makes these ideas concrete by constructing the minimal sufficient states—“causal states”—needed for optimal prediction. These states are defined such that all histories that lead to the same causal state yield identical conditional distributions over futures. For a simple Markov chain, the causal states coincide with the chain’s observable states, and the process has no additional hidden memory. For more structured processes, the causal states may be numerous, high-dimensional, and history-dependent. The amount of information needed to specify which causal state the system is currently in is called the statistical complexity. High statistical complexity indicates that a process has rich internal memory, even if its observable behavior looks deceptively simple. When applied to neural recordings or behavioral sequences, this framework reveals that the underlying effective state space of cognition is far more elaborate than what a Markovian description allows.

Information theory also provides tools for quantifying how memory is distributed across scales rather than localized in a single buffer. Multiscale entropy and related measures examine how predictability changes as one coarse-grains the temporal resolution of a signal. Neural time series often exhibit power-law scaling and 1/f-like spectra, indicating that fluctuations at many timescales contribute to behavior. Markov models, by contrast, impose a characteristic timescale defined by their transition structure. When empirical data reveal no such privileged scale—when past influences decay slowly over many orders of magnitude—this points to an inherently non-markovian regime. The system behaves as if it possesses a continuum of partially overlapping memory traces, with no clean cut between “relevant” and “irrelevant” history.

These considerations highlight why attempts to rescue Markovian descriptions by simply “expanding the state space” are only partially successful. In principle, one can always transform a finite-order non-Markov process into a higher-dimensional Markov process by embedding enough past time steps into the state. However, if the relevant dependencies span unbounded or widely distributed timescales, this embedding would require an effectively infinite-dimensional state. Information-theoretic quantities such as statistical complexity can diverge under these conditions, signaling that no finite Markov model can capture the process without substantial loss of predictive information. For cognition, this means that while local Markov approximations may be useful within restricted contexts, they cannot serve as complete explanations of how the mind integrates experience over time.

There is a further conceptual issue: Markov models assume that all relevant information is internalized into a well-defined state at each moment. But information theory allows us to consider extended systems in which informational dependencies span brain, body, and environment. For example, the layout of a familiar workspace, notes on a piece of paper, or the presence of a smartphone can all function as externalized memory that reduces uncertainty about future actions and outcomes. When these external resources are stable and systematically coupled to internal processing, they form part of an extended information-processing loop. From an information-theoretic standpoint, the effective state of the cognitive system then includes these external variables, and the overall process becomes deeply non-markovian when viewed from the narrower perspective of the brain alone.

The tension between non-markovian cognition and physical causality sometimes leads to speculative discussions of retrocausality or “influences from the future.” Information theory offers a more sober framing. In a temporally extended, non-Markov system, the present can carry information about both the past and the future, in the sense that knowing the current state can improve our inferences about what has already happened and what will happen next. This bidirectional informational constraint does not entail that future events literally cause present states; rather, it reflects the way that underlying dynamical and probabilistic structure imposes correlations across time. For a sufficiently sophisticated observer, updating beliefs about the future can lead to revised inferences about the past, and vice versa, without any violation of causal order. The apparent “backwards influence” is an artifact of inference in a complex temporal web, not a sign that time is running in reverse.

In light of these analyses, the limits of markovian minds can be summarized in strictly information-theoretic terms. A Markov agent treats the present state as a complete summary of all information relevant for prediction and control. Whenever the environment, the body, or the neural substrate itself exhibits longer-range dependencies, this assumption guarantees an information bottleneck: potentially useful bits about extended history are irrevocably discarded. Non-markovian architectures, by contrast, preserve and exploit structured dependencies using distributed memory, recurrent connectivity, and temporally deep generative models. Information theory does not merely catalog these differences; it shows that, under realistic conditions of environmental complexity and resource constraints, non-markovian strategies are not optional embellishments but prerequisites for near-optimal prediction, learning, and action.

Empirical evidence for non-markovian dynamics in cognition

Empirical support for non-markovian dynamics in cognition comes from multiple levels of analysis: overt behavior across time, neural dynamics measured with electrophysiology and imaging, and fine-grained studies of memory, learning, and decision-making. Across these domains, the common finding is that the influence of past events on present processing cannot be fully summarized by a finite, short-lived internal state. Instead, history leaves structured, path-dependent traces that continue to shape cognition over extended intervals.

Consider first the classic laboratory paradigms used to probe human memory. Serial position effects in free recall, such as primacy and recency, show that items encountered early and late in a list enjoy special advantages, implying that the impact of each item on recall depends on its position within a broader temporal sequence. More tellingly, contiguity effects reveal that when people recall one item, the next item retrieved is disproportionately likely to have appeared nearby in the original list. This “temporal clustering” persists even when list items are semantically unrelated, indicating that the memory system encodes not only what was presented, but also the specific temporal order in which it appeared. Models that treat recall as a Markovian process, where each retrieval depends solely on the currently active memory trace, systematically underestimate this rich temporal structure.

Studies of episodic memory further highlight non-markovian dependencies. When participants view or listen to continuous narratives—films, stories, or realistic event sequences—subsequent recall tends to preserve not just individual elements but higher-order temporal organization. People recall scenes in roughly the original order, group related segments into “events,” and omit or compress intermediate details. Neuroimaging reveals that transitions between such remembered events correspond to large-scale reconfigurations of brain activity, particularly in the hippocampus and default mode network. The brain appears to maintain a temporally extended representation of narrative structure that guides both encoding and retrieval. This representation cannot be reduced to a step-by-step Markov chain; it behaves more like a latent event model in which the meaning and memorability of each moment depend on its role within a longer trajectory.

Neurophysiological studies in animals provide even more direct evidence that the brain carries detailed information about past and potential future paths in ways that violate Markov assumptions. Hippocampal place cells in rodents fire in location-specific patterns during navigation, but they also exhibit replay and preplay: compressed sequences of firing that recapitulate past trajectories or simulate possible future paths while the animal is resting or pausing. These sequences can span many spatial locations and seconds of real-world time, linking nonadjacent states into coherent trajectories. Importantly, which paths are replayed depends on prior experience, rewards, and task demands, demonstrating that the hippocampal code embodies a history-dependent, goal-sensitive representation of space and action. A purely Markovian model, in which the firing pattern at each moment depends only on the animal’s current position, would miss this internally generated temporal structure.

Complementary evidence comes from studies of sequence learning and statistical learning, where participants are exposed to streams of stimuli with hidden patterns. In artificial grammar learning tasks, for example, subjects view letter or tone sequences generated by rules they are never told explicitly. Over time, they become sensitive to which sequences are “grammatical,” and this sensitivity often extends to long-range dependencies that cannot be captured by first-order transition probabilities between adjacent elements. Similarly, in motor sequence learning, such as pressing keys in response to a repeating pattern of cues, performance improvements reflect knowledge of higher-order structure: people anticipate multi-step regularities, not just the next immediate element. Behavioral performance and reaction times show that the system retains detailed information about sequences of states, not merely the most recent one.

Language processing offers especially fine-grained tests of temporally deep structure. Eye-tracking studies reveal that pronoun resolution, syntactic disambiguation, and discourse coherence judgments depend on information encountered many words, sentences, or even paragraphs earlier. For instance, the ease with which a reader interprets a pronoun like “she” depends on the distribution of potential antecedents across the preceding text, as well as on broader discourse goals. Event-related potential (ERP) components such as the N400 and P600 are modulated not only by the immediately preceding word but by global context, genre expectations, and prior knowledge established earlier in the narrative. These neural signatures indicate that the brain maintains and updates long-range priors about linguistic structure and meaning, enabling prediction that reaches far beyond what a local Markov model would allow.

Decision-making research likewise uncovers robust non-markovian patterns. In reinforcement learning paradigms, participants often exhibit history-dependent biases like win-stay/lose-shift strategies, perseveration, and sensitivity to streaks of success or failure. More sophisticated tasks, such as multi-step planning or “two-step” decision tasks, show that people build internal models of the task’s structure that track not only current states and rewards but the contingencies linking sequences of actions to distal outcomes. Neuroimaging reveals that prefrontal and hippocampal activity patterns during planning reflect simulations of multi-step paths, suggesting that choice behavior is driven by mental trajectories through possible futures, rather than by a simple mapping from current state to action.

Studies of habit formation and extinction in both animals and humans also challenge Markovian assumptions. When an action has been repeatedly associated with a reward, organisms often continue performing it even after the reward is removed, a phenomenon known as persistence or habit. The rate at which habits form and are extinguished depends on extended training history: the schedule of previous reinforcements, the variability of outcomes, and the spacing of trials. These effects cannot be predicted solely from the last few trials; instead, the entire reward history shapes internal valuations and response propensities. Neural recordings from striatum and orbitofrontal cortex reveal signals that integrate reward prediction errors over long windows, supporting a non-markovian accumulation of evidence about action–outcome contingencies.

Neural dynamics during resting state and task engagement further corroborate temporally extended dependencies. Spontaneous fluctuations in fMRI signals and electrophysiological recordings exhibit 1/f-like power spectra and long-range temporal correlations, indicating that activity at any given moment is statistically related to activity many seconds or minutes earlier. Measures such as detrended fluctuation analysis and multiscale entropy demonstrate that these correlations persist across multiple temporal scales, with no single characteristic timescale dominating. This pattern is incompatible with simple low-order Markov models, which produce exponential decay of correlations. Instead, the data suggest that the brain operates as a multiscale, non-markovian system in which past states leave graded, slowly decaying influences on current processing.

Task-based neuroimaging supports a similar picture. In working memory tasks, for example, prefrontal and parietal areas maintain elevated, pattern-specific activity during delay periods, preserving information about stimuli that are no longer present. Yet the dynamics within these regions are not static: subtle drifts and oscillatory modulations reflect interactions between maintained content and ongoing sensory, motor, and internal signals. Decoding analyses show that patterns of activity at later time points can be better predicted when models incorporate multiple previous time points, rather than a single instantaneous snapshot. This implies that the effective neural state is a function of a trajectory through neural state space, not just of the current input and last step.

More sophisticated analyses of neural recordings use tools from dynamical systems theory to reconstruct trajectories in latent spaces. In motor cortex, for example, complex movements can be described as low-dimensional trajectories that unfold smoothly over time. These trajectories depend on preparatory states that begin well before movement onset and can be influenced by previous trials, errors, and rewards. Perturbations introduced at one moment can have consequences that unfold over extended periods, revealing that the system’s future evolution depends on its entire recent path through state space. The same is true in cognitive control tasks: prefrontal cortex trajectories reflect not only current stimuli but accumulated conflict, errors, and task rules that have been in effect over many trials.

Empirical work on context and state-dependent processing in perception also reveals non-markovian characteristics. Adaptation effects—such as tilt aftereffects in vision or pitch adaptation in audition—demonstrate that recent stimulus history shapes sensitivity to current inputs. These effects cannot be fully described by a single, fast-decaying adaptation variable; instead, multiple overlapping adaptation processes operate over different timescales, from milliseconds to minutes. Perceptual hysteresis, where ambiguous stimuli (like a Necker cube or bistable motion) tend to be interpreted consistently with their recent perceptual history, points to slow, cumulative biases in the inference process. The resulting patterns of switching and persistence are better captured by models that include temporally deep priors and history-dependent attractor dynamics than by memoryless decision rules.

Emotion and mood provide another rich source of evidence. Affective states show inertia: mood at one moment strongly depends on mood at earlier moments, and the decay of this dependence can be slow and nonlinear. Time-series analyses of self-reported affect reveal autocorrelations stretching across hours or days, and these dynamics differ systematically between individuals and clinical populations. For instance, depression is associated with sluggish recovery from negative events and prolonged dwelling on negative thoughts, suggesting that affective processing has altered temporal integration properties. Neuroimaging studies align with this, showing that amygdala, insula, and prefrontal regions display prolonged responses to emotional stimuli and sustained coupling within large-scale networks. These persistent responses mean that current appraisals and reactions are shaped by a temporally thick backdrop of prior emotional context.

Empirical phenomena often labeled as “intuition” or “gut feeling” can also be interpreted through non-markovian lenses. In complex tasks such as chess, medical diagnosis, or financial decision-making, experts appear to make rapid, accurate judgments based on limited explicit information. Process-tracing studies suggest that these judgments rely on rich internal models formed over long experience, where subtle cues in the present trigger associations to extensive histories of similar patterns and outcomes. Neural signatures in experts show rapid activation of distributed networks, including memory and reward circuits, upon encountering familiar configurations. The present cue thus taps into an implicitly stored, high-dimensional history that shapes immediate response, again undermining the idea that current behavior can be fully explained by local, Markovian states.

Experimental manipulations that alter temporal structure without changing immediate stimuli provide strong tests of non-markovian processing. In temporal context experiments, identical target stimuli are embedded in sequences with different statistical properties: for example, in one condition, targets follow predictable rhythms; in another, they appear irregularly. Participants’ detection accuracy and timing judgments depend on the broader temporal context, even when the local interval before each target is the same. Neural measures show that oscillatory phase alignment and anticipatory activity in sensory cortices track the higher-order temporal structure, implying that prediction and attention are guided by accumulated information about patterns across many trials.

Another line of evidence comes from cross-session and cross-day effects in learning. Skills practiced on one day often show offline improvements after sleep, and performance on related tasks can be facilitated or interfered with by prior training, even when those tasks are separated in time and differ in surface features. Sleep-dependent consolidation studies in both humans and animals reveal that hippocampal and cortical replay during sleep recapitulates patterns from prior wakefulness, selectively strengthening some memories while weakening others. The influence of these consolidation processes on later behavior demonstrates that the system’s effective learning state is not fixed at the end of each session, but continues to evolve based on internal reprocessing over extended intervals. This intrinsic, offline history-dependence cannot be accommodated by Markovian models that treat post-learning periods as inert.

Clinical and developmental data underscore that cognitive trajectories are deeply path-dependent. Early experiences, including stress, enrichment, deprivation, and social interaction, have long-lasting effects on attention, emotion regulation, and executive function. Longitudinal studies show that small differences in early environments can lead to diverging developmental paths, with effects that are not easily reversible by later interventions. At the neural level, these histories are reflected in stable differences in connectivity patterns, neuromodulatory baselines, and gene expression profiles. The resulting cognitive and behavioral profiles at any given age are thus the cumulative outcome of extended, branching histories—precisely the sort of non-markovian dependence that simple state-based models fail to capture.

Implications for consciousness, agency, and free will

Non-markovian models of mind alter how consciousness is situated in time. If experience is not generated solely from a thin, instantaneous state, but from trajectories and temporally deep generative models, then subjective awareness is always occupied by more than the present. Perception arrives already filtered through accumulated memory, slowly evolving expectations, and latent narratives about where things are coming from and where they are going. The phenomenal “now” becomes a moving window carved out of a much thicker process: a surface where prior experiences, implicit priors, and ongoing prediction about the near future all converge. In this sense, consciousness is less like a series of static snapshots and more like the visible crest of a wave whose form depends on long histories of motion beneath the surface.

This temporally thick picture helps explain why consciousness feels continuous despite the brain’s discrete spikes and oscillations. Non-markovian neural dynamics allow information about past states to remain functionally active, so that a moment of awareness inherits structure from what has just occurred and what is anticipated. Consider listening to a melody: the present note is experienced not as a bare frequency but as continuation, confirmation, or violation of an ongoing pattern. The aesthetic and emotional quality of the note depends on remembered motifs and expected resolutions. A Markovian account, tied to the current note alone, cannot capture this phenomenology. By contrast, a temporally deep generative model provides a natural bridge between subjective continuity and the brain’s extended integration of past and predicted future states.

Agency, in this framework, becomes a matter of how an organism harnesses its temporal depth to shape outcomes. A purely Markovian agent reacts based on the current state; a non-markovian agent draws upon accumulated evidence, learned structure, and anticipatory simulations to select actions. The capacity for planning illustrates this difference. To plan is to construct and evaluate counterfactual futures, then let those imagined trajectories influence present choice. This presupposes that the system can maintain and compare multiple possible futures, retain them across delays, and update them as new information arrives. Such temporally extended operations make agency more than immediate reaction; they imbue action with a sense of direction, commitment, and responsibility rooted in an internal history of deliberation.

On a predictive processing and bayesian brain view, agency is tightly bound to control of prediction error over time. Organisms act to reduce anticipated discrepancies between their preferred states and expected trajectories of the world. When internal models are sufficiently rich, the system can consider diverse policies—alternative sequences of action—and evaluate their long-run consequences. The selected policy is not determined by the present state alone but by stored knowledge of what similar choices have led to in the past and by expected future opportunities and risks. This means that at the moment of action, many moments—past learning episodes, prior commitments, and projected futures—are co-present in the causal background. Responsibility for action, on this picture, attaches not to an isolated instant but to the extended, path-dependent process that produced and endorsed the policy.

Non-markovian architecture also reshapes familiar debates about free will. Traditional incompatibilist worries often rely on a picture in which the state of the world at a time, together with deterministic laws, fixes the next state. If the mental state at t is just one node in a Markov chain, fully determined by the state at t−1, then freedom appears illusory. However, when the effective state relevant for action includes history-dependent variables—long-term memories, slowly changing motivational landscapes, traits, and internalized norms—the causal story is less like a single-step push and more like the unfolding of a complex trajectory. Determinism may still hold at the physical level, but the psychologically salient causes of action are distributed over extended periods of learning and self-organization. The agent is, in part, the pattern that emerges from these temporally thick processes.

This suggests a compatibilist reinterpretation of freedom in explicitly temporal terms. Rather than asking whether a momentary state could have been otherwise, one can ask whether the long-run organization of the system allowed for meaningful sensitivity to reasons, counterfactual considerations, and new evidence. A non-markovian mind that updates its priors, revises its goals, and reshapes its own dispositions over time can exhibit a form of diachronic autonomy, even if each step is locally determined. Freedom, on this view, consists in being the kind of process whose future trajectory depends delicately on reasons and information integrated across extended histories, not in being exempt from causation. The richer and more revisable the internal temporal structure, the more room there is for rational self-governance.

Temporally deep models also clarify the role of memory in agency. Episodic and semantic memories are not inert archives but active constraints on what actions feel available, appropriate, or salient. Remembering past successes can expand the space of considered options; recalling failures or harms can narrow it. Because memory retrieval itself is reconstructive and influenced by current goals and expectations, acts of remembering can modify the very evidence base on which future decisions rest. This feedback loop means that individuals partly author their own causal history by how they attend to and reinterpret the past. Agency thus includes not just choosing actions in the moment but managing the narratives and memories through which those choices are framed.

The phenomenon often called “memory of the future” adds another dimension. When an agent repeatedly simulates specific future scenarios—rehearsing a difficult conversation, catastrophizing about a possible loss, or savoring a hoped-for success—these imagined episodes can acquire the same motivational force as recollected events. They bias attention, shape affect, and influence policy selection. From the standpoint of the generative model, frequently rehearsed futures become entrenched priors: they are treated as highly probable or especially important potential outcomes. In this way, the future exerts an indirect influence on present action, not through literal retrocausality, but through the way anticipated possibilities are encoded and re-encoded as if they were memories that must be honored or avoided.

Subjectively, this can fuel the intuition that the future is “pulling” us toward certain outcomes. A person driven by a long-term project may experience present sacrifices as demanded by a future state of affairs—the finished book, the recovered health, the secure family. In mechanistic terms, what is happening is that temporally extended goals have been installed as high-level priors that shape lower-level predictions and policy evaluation. The agent’s conscious sense of being drawn forward reflects the dominance of these temporally distant constraints within the non-markovian control architecture. Free will, in this lived sense, is experienced as aligning current actions with strongly represented, temporally deep goals, even though the causal direction still runs from past learning and present simulation to future behavior.

Agency is also bound up with self-modeling, and self-models are inherently non-markovian. A self is not merely a pointlike observer but a temporally extended entity with a past and a projected future. To experience oneself as an agent is to integrate memories of prior actions, attributions of responsibility, social feedback, and expectations about one’s own behavior into a coherent, if evolving, pattern. This pattern is implemented by neural dynamics that maintain cross-temporal associations: the same person who failed yesterday can resolve to improve tomorrow; the same person who made a promise earlier can feel bound by it now. Because the self-model tracks continuity across time, it can anchor commitments, regrets, and plans. These phenomena are difficult to make sense of under strictly Markovian descriptions that privilege only immediate states.

From an ethical standpoint, path-dependence complicates how we assign praise and blame. If present capacities for control, impulse regulation, or empathy are the result of long developmental trajectories—shaped by early environment, social context, and stochastic events—then responsibility appears to be spread across many earlier conditions. Non-markovian minds carry deep imprints of circumstances for which the individual had little or no control. At the same time, because agents can, over long periods, modify their own generative models by seeking therapy, education, or new environments, there remains room for self-directed change. Ethical and legal notions of responsibility may need to balance these insights: actions at a given time are both the downstream product of prior influences and potential turning points in the ongoing reconfiguration of the agent’s temporal architecture.

These considerations have direct implications for the design of artificial agents. Many current AI systems used in sensitive domains—recommendation engines, credit scoring, predictive policing—are effectively Markovian at the decision layer, conditioning on a snapshot of available data. Introducing genuinely non-markovian architectures, with long-range memory of prior outputs and user interactions, could enhance performance but also magnify path-dependence. Recommendations could create stronger feedback loops, and early model errors could imprint themselves on the system’s future behavior in hard-to-reverse ways. Understanding agency in non-markovian terms therefore matters not only for human free will but for assessing the autonomy and responsibility of artificial systems that learn and act over extended timescales.

Non-markovian cognition also reframes worries that physical determinism leaves no room for genuine choice. When the mind is modeled as a temporally deep inference engine rather than a memoryless transition system, the relevant question shifts from “Could the current microstate have produced a different action?” to “How does the long-run organization of this process respond to reasons, evidence, and anticipated futures?” Deterministic neural dynamics can still implement flexible, history-sensitive choice architectures whose behavior varies systematically with argument, information, and reflection. In that setting, talk of freedom and responsibility picks out properties of the extended pattern—the way priors can be revised, goals reprioritized, and habits reshaped—not the metaphysical status of any single, instantaneous state.

The possibility that consciousness itself depends on specific non-markovian features of neural dynamics cannot be ruled out. Some theories propose that conscious states require integration of information over particular temporal windows, or the maintenance of metastable patterns that bind multiple modalities and timepoints into unified episodes. If so, then disrupting temporal integration—by fragmenting experience into uncorrelated segments or severely limiting access to past context—might degrade or abolish consciousness even if instantaneous processing remained intact. This would mean that subjective awareness is not merely a byproduct of local computations but an emergent property of processes that carry, transform, and compare information across time. In that case, the very existence of conscious agents with a sense of self and freedom would depend on the non-markovian fabric of their cognitive lives.