In predictive processing, the organism is understood as a continually updating generative model that anticipates its own future states as well as the sensory consequences of its actions. Temporal self-modeling refers to the way this generative model constructs, maintains, and revises a representation of the self extended through time, rather than merely tracking instantaneous bodily or perceptual states. Instead of a static entity, the self is treated as an evolving probabilistic hypothesis about who one is, what one is doing, and what one will likely experience or become. This hypothesis is encoded in structured expectations about trajectories: sequences of states and actions that unfold over multiple time scales. Every moment of experience is thus situated within a predictive context in which the present self is implicitly tethered to remembered past selves and anticipated future selves.
Under the predictive processing framework, perception is not a passive registration of inputs but an active inference process in which the brain continuously tests predictions against incoming sensory signals. The so-called bayesian brain uses hierarchically organized priors—learned expectations about hidden causes in the world and body—to generate top-down predictions. The self emerges within this architecture as a special class of hidden cause: a persistent, structured source that explains the coherence of bodily sensations, internal states, and action patterns over time. Temporal self-modeling occurs as the system infers a best-fitting narrative that links past, present, and prospective states into a single, minimally surprising trajectory. This narrative is not explicitly verbal; it is instantiated as probabilistic structure embedded in neural dynamics that anticipate how the organism will change as it acts and is acted upon.
Because prediction in this setting is always temporally deep, the self-model is inherently future-oriented. The organism does not merely estimate its current physiological and sensorimotor condition; it forecasts likely outcomes of possible actions and adjusts behavior to minimize expected prediction error over time. Temporal self-modeling therefore involves assigning probabilities to alternative self-trajectories and selecting those that are expected to maintain internal coherence and viability. The generative model encodes conditional dependencies linking earlier and later states, such that what counts as “me” now is partly determined by what I expect to be doing and feeling moments, hours, or even years from now. This forward-looking structure transforms the self from a snapshot into a dynamically stabilized pattern across time.
Prediction error minimization provides the basic engine for sculpting these temporally extended self-hypotheses. When there is a mismatch between expected and observed states—whether bodily, affective, or social—the system can respond in two principal ways: by updating its self-model to better fit the surprising data, or by acting to change the data so that it conforms more closely to existing expectations. Temporal self-modeling arises from the interplay of these two strategies across time. By updating, the organism revises who it takes itself to be (for example, adjusting beliefs about its abilities or vulnerabilities). By acting, it selectively samples environments and situations that stabilize its preferred self-predictions (for example, seeking contexts where its skills or traits are reaffirmed). Over repeated cycles, a relatively stable, temporally extended sense of self crystallizes from this continuous negotiation between expectation and evidence.
A crucial feature of temporal self-modeling in predictive processing is the hierarchical organization of time scales. High-level priors encode slowly changing aspects of the self—such as personality traits, enduring values, and long-term social roles—while lower-level priors track rapidly fluctuating bodily states, momentary intentions, and local perceptions. This hierarchy enables the system to maintain a coherent long-range self-narrative even as immediate experiences shift from moment to moment. Rapid fluctuations at lower levels are interpreted in light of more stable higher-level expectations, and prediction errors are selectively propagated upward only when they are strong or persistent enough to warrant revising more abstract aspects of the self. In this way, the generative model maintains temporal consistency without becoming rigid or insensitive to genuine change.
Within this architecture, the feeling of being a continuous subject can be understood as the experiential correlate of successful temporal self-modeling. When predictive processing proceeds smoothly, with manageable prediction errors and effective updates, consciousness presents an impression of seamless flow from one moment to the next. The self is not explicitly represented as a static object but as the implicit point of reference from which predictions are generated and evaluated. Disruptions to temporal self-modeling—such as sudden, large deviations from expected bodily states, social identities, or emotional patterns—can fragment this sense of continuity, leading to experiences of alienation from one’s own thoughts or actions. Thus, the phenomenology of selfhood reflects the stability and precision of temporally extended priors that knit together diverse experiences into a unified stream.
Action and temporal self-modeling are inseparable under predictive processing. To minimize long-run prediction error, the system must anticipate the consequences of prospective actions and adjust its behavior accordingly. This means the self-model includes not just estimates of what one is and has been, but of what one can do and is likely to do. Anticipated actions and their expected sensory outcomes are woven into the temporal structure of self-representation, so that the self is partly defined by a repertoire of possible futures. When these expected futures are reliable and richly specified, the agent experiences a robust sense of agency and continuity. When they are uncertain, truncated, or internally inconsistent, the temporal self-model becomes fragile, and the experiential sense of being a unified, temporally extended self may waver.
Temporal self-modeling also depends on how the generative model balances precision across time. The system must decide how much confidence to place in priors about long-term identity versus short-term fluctuations. If long-range priors are assigned excessively high precision, the self-model may become resistant to updating, preserving a rigid or unrealistic picture of who one is despite accumulating contradictory evidence. If, on the other hand, low-level, rapidly changing signals are granted too much precision, the self may appear unstable, excessively reactive to transient states and contexts. Effective temporal self-modeling involves a calibrated allocation of precision that allows the self to be both resilient and adaptable, maintaining overall coherence while remaining sensitive to genuine change in circumstances or capacities.
Generative entanglement across past and future selves
When the generative model is temporally deep, the self is not simply an extrapolation of the past but an active entanglement of past and future hypotheses. The bayesian brain does not first infer a purely retrospective identity and then, as a separate step, project this identity into the future. Instead, it jointly optimizes beliefs about earlier, present, and later self-states under a single, temporally extended probability distribution. In this distribution, expectations about the future constrain how the past is encoded, just as memories of the past constrain which futures are deemed plausible or desirable. This bidirectional constraint structure produces what can be called generative entanglement: the fact that neither “who I was” nor “who I will be” is independently determined, but each is inferred in light of the other under the imperative to minimize prediction error across an extended temporal horizon.
From this perspective, temporal inference about the self is formally akin to smoothing in Bayesian time-series models, where beliefs about intermediate states are shaped by both prior and subsequent observations. Applied to lived experience, the implication is that the generative model continually re-estimates earlier self-states in light of newly inferred trajectories. A present failure or success, for example, does not merely update expectations about what one will do next; it can retroactively alter how previous efforts are interpreted, recasting them as early signs of incompetence, perseverance, or latent ability. This is not retrocausality in the physical sense, but a form of epistemic retrocausality: future-oriented predictions feed back to reshape the probabilistic structure of the remembered past. The self becomes a time-symmetric inference problem in which the entire life trajectory is repeatedly re-fit as new data arrive.
This entanglement is especially evident in the way counterfactual futures reorganize autobiographical memory. The generative model routinely entertains alternative paths the self could take—different careers, relationships, or ways of being—and evaluates them in terms of expected prediction error and affective cost. As these counterfactual futures are explored, they selectively amplify or suppress particular memories that support or undermine their plausibility. A contemplated identity as a “creative person,” for instance, can prompt a reinterpretation of scattered childhood episodes—drawing, storytelling, improvisation—as evidential anchors for that imagined trajectory. Conversely, if a future is deemed untenable or too costly, the model may downplay or reorganize past events that once seemed to support it. Past and future selves are thus co-constructed as mutually reinforcing probabilistic narratives rather than independently stored segments of experience.
Generative entanglement also operates at the level of affective and evaluative structure. The generative model does not encode the self as a neutral sequence of events, but as a pattern of value-laden states with expected hedonic, social, and bodily consequences. Future-valuations—projections of what will feel good, bad, meaningful, or shameful—propagate backward through the model and color the subjective salience of earlier episodes. Experiences that align with anticipated valued futures are assigned higher precision; they are encoded more stably and recalled more readily. Experiences that conflict with anticipated futures may be rendered noisier, discounted, or walled off. In this way, the affective landscape of possible futures functions like a set of gravitational wells that pull the remembered past into configurations that stabilize preferred self-trajectories and dampen trajectories that would be intolerable or disorganizing.
At the level of moment-to-moment inference, entanglement manifests as an ongoing negotiation between short-range expectations and long-range identity constraints. On short time scales, the system predicts immediate sensory and motor states—muscle tensions, speech sounds, gaze shifts—guided by local context. On longer time scales, it maintains abstract expectations about what “someone like me” typically does or should do in the current situation. When local predictions deviate from these higher-level identity priors, the system can modify immediate behavior to bring it in line with the expected self, or relax and update identity priors to better accommodate the emergent pattern of action. Over time, this two-way adjustment causes patterns of behavior, and hence of experienced agency, to crystallize along trajectories that are jointly shaped by remembered tendencies and anticipated futures. The self is thereby enacted as a compromise solution to a temporally global error-minimization problem.
Phenomenologically, this can be felt as a subtle but pervasive forward-directedness in consciousness. Even when attention appears focused on recollection or on the present scene, experience is structured by an undercurrent of “where this is going” and “what this says about me.” A conversation, a bodily sensation, or a fleeting mood is not merely perceived; it is tacitly situated within a space of possible continuations that carry implications for future roles, relationships, and self-evaluations. This orientation means that the lived sense of being a self is always already stretched between what has ostensibly happened and what is expected to follow. When predictive processing is functioning coherently, this stretching is experienced as an ongoing narrative arc in which past, present, and future hang together. When coherence breaks down—for instance, under conditions of radical uncertainty or conflicting long-term expectations—the entanglement can become experienced as tension, indecision, or fragmentation in one’s sense of who one is.
Generative entanglement is not merely a byproduct of limited memory or imagination; it is a computational resource that allows the organism to leverage evidence across time. By letting future-directed expectations refine beliefs about what sort of past one must have had, the model compresses the vast space of possible life histories into a smaller set of self-consistent trajectories. These trajectories can then guide action and perception in a way that is both efficient and robust to noise. For example, adopting a long-term expectation of being a dependable friend constrains which past social episodes are treated as diagnostic and which current behaviors are seen as self-congruent. This reduces ambiguity in social situations and streamlines decision-making, at the cost of potentially distorting or omitting experiences that would unsettle that identity. The entanglement thus trades representational fidelity against pragmatic stability, favoring trajectories that support coherent and effective engagement with the world.
This same structure implies that interventions at one temporal point can propagate widely across the self-model. A seemingly small adjustment in long-term expectations—such as revising one’s sense of future competence or belonging—can ripple backward to reframe entire chapters of autobiographical memory and forward to reshape patterns of attention, exploration, and risk-taking. Likewise, a single unexpected event in the present can force a reconfiguration of both remembered and anticipated selves if it generates sufficiently large and persistent prediction errors. In these cascades of updating, what remains “the same person” is not a fixed inner core but the continuity of the generative process itself: the ongoing attempt to maintain a globally coherent pattern of expectations across time. The self persists as the dynamically conserved structure of this time-entangled generative model, rather than as any particular content it momentarily endorses about the past or the future.
Hierarchical time scales in self-representation
Self-representation in predictive processing is distributed across a hierarchy of temporal scales, each encoding different aspects of the organism’s ongoing existence. At the shortest scales—tens to hundreds of milliseconds—the generative model tracks fast-changing sensorimotor contingencies: proprioceptive feedback from limbs, fluctuations in heart rate, moment-to-moment visual and auditory input, and the microdynamics of speech and gaze. These rapidly updated states implement the fine-grained sense of “what I am currently doing” and “how my body is configured right now.” At intermediate scales—seconds to minutes—the model integrates these transient events into recognizable episodes: engaging in a conversation, preparing a meal, walking to a destination. At still longer scales—days, months, years—the system maintains more abstract patterns, such as social roles, enduring projects, and life themes. The self emerges not from any single level but from the coordination of these nested time scales, each constraining and being constrained by the others.
Within this hierarchy, higher levels encode slowly varying priors that stabilize identity-relevant expectations. These might include beliefs such as “I am a reliable friend,” “I am physically fragile,” or “I am the kind of person who avoids conflict.” Such priors change only gradually because they integrate evidence over long spans of time and many contexts. Lower levels encode more volatile expectations: “my heart is racing,” “my voice is shaking,” “I am about to say something risky.” The predictive processing architecture allows faster levels to adjust quickly to local conditions while remaining tethered to slower identity constraints that provide a sense of overall coherence. As long as lower-level prediction errors can be absorbed without challenging high-level priors, the agent experiences continuity: unusual bodily states or emotional reactions are interpreted as local deviations rather than threats to who one takes oneself to be.
The interaction across time scales is mediated by precision-weighting. Higher levels determine how much weight to assign to discrepancies at lower levels when deciding whether to revise slow-changing self-beliefs. A transient spike of anxiety before public speaking, for instance, may be treated as noise if high-level priors about being “generally competent and socially accepted” are strong and precise. The momentary discomfort is then explained away as situational, and the broader self-model remains intact. If, however, similar episodes occur repeatedly and prediction errors accumulate, their effective precision increases. Eventually, they may propagate upward to revise more abstract expectations, prompting a shift toward a self-concept such as “I am socially anxious” or “I am not good at public speaking.” In this way, repeated patterns of low-level surprise can gradually reshape higher-level identity priors, while those priors simultaneously filter which surprises are allowed to matter.
This hierarchical organization can be thought of as a cascade of temporal smoothing operations. At each level, the generative model compresses faster fluctuations into slower, more stable summaries that serve as priors for the next level down. Momentary affective states are smoothed into moods, moods into personality traits, individual decisions into habits, and habits into enduring character. From the inside, this smoothing process is felt as a continuity of character across changing circumstances: one recognizes oneself as “the same person” even as particular feelings, thoughts, and behaviors vary. The sense of an enduring self thus depends on the success of these smoothing operations in filtering out noise and preserving patterns that are statistically robust over time. When smoothing fails—because signals are too chaotic, or precision is misallocated—the self can feel fragmented or volatile, as though one is constantly becoming someone new without a stable throughline.
Hierarchical time scales also structure agency. At short scales, the system generates detailed motor predictions about specific movements and their immediate consequences: reaching for a glass, articulating a sentence, shifting posture. At intermediate scales, it encodes policies that implement recognizable actions: giving a talk, resolving a disagreement, completing a task. At the longest scales, it represents projects and commitments that can span years: raising a child, pursuing a career, maintaining a moral or political stance. These levels are coordinated such that fine-grained motor control is continually steered by medium-range policies, which in turn are selected and evaluated against long-range projects and values. The feeling of acting “in character” arises when local actions and short-term policies align with these long-term expectations; the feeling of acting “out of character” signals a mismatch between short and long temporal scales of self-representation.
An important consequence of this multi-scale architecture is that different components of the self-model can reorganize at different speeds. Rapid contextual shifts—moving to a new city, entering an unfamiliar social environment—can produce immediate changes in behavior and affect at lower levels without instantly revising slow-changing identity priors. The organism effectively runs a temporary “context mode,” adjusting lower-level predictions to handle novel demands while preserving core assumptions about who it is. If the new context persists and remains mismatched with existing priors, the system faces a choice: either reshape the environment and one’s niche to better match entrenched self-expectations, or gradually update those expectations to accommodate the new pattern of evidence. Identity change thus often proceeds as a lagging reconfiguration of high-level priors in response to sustained pressure from mismatched lower-level dynamics.
This temporal staggering helps explain why self-representation can exhibit both inertia and sudden shifts. Identity priors are conservative because they are backed by large amounts of accumulated experience, but they can reorganize relatively quickly when confronted with overwhelmingly precise, persistent prediction errors that cannot be absorbed by local reinterpretation. A single traumatic event, for example, can reweight precision at multiple levels, leading to rapid wholesale changes in core expectations about safety, trust, or competence. Conversely, subtle but consistent feedback—small successes or failures repeated over months or years—can slowly tilt high-level priors without any single dramatic episode. The same architecture thus accounts for both apparent stability of character and gradual or abrupt transformations in how one understands oneself.
Differentiated temporal scales also partition the phenomenology of consciousness. Very fast predictive loops underpin the pre-reflective sense of embodied presence: the feeling of occupying a body that moves fluidly and reliably responds to intentions. Intermediate loops organize the stream of experience into meaningful episodes, structuring consciousness into units like “an argument,” “a walk,” or “an evening with friends.” Slow loops support autobiographical continuity, sustaining an implicit background sense of “my life” as an ongoing story with a past and possible futures. When these scales are well-coordinated, consciousness presents as a layered but unified field in which momentary sensations, current activities, and long-term concerns all feel like expressions of a single, temporally extended self. Discoordination—such as when immediate bodily signals seem disconnected from one’s life narrative—can produce experiences of derealization, depersonalization, or a sense of watching one’s life from the outside.
The hierarchy of time scales also offers a way to understand role-based and context-sensitive selves without positing multiple independent selves. High-level identity priors may themselves be stratified: some encode extremely slow, cross-context traits (“I value honesty,” “I care for my children”), while others encode roles with intermediate temporal scope and situational specificity (“I am a manager at work,” “I am a student in this course”). Each role-prior brings with it expectations about appropriate actions, emotional responses, and patterns of attention that unfold over characteristic intervals—meetings, semesters, projects, seasons. As situations change, the generative model selectively activates different subsets of these priors, constraining the lower-level dynamics accordingly. The experience of “shifting into” a professional, parental, or intimate mode reflects a reconfiguration of which mid-level time-scale structures are currently dominating prediction and error minimization.
Because these role-based priors inhabit intermediate time scales, they can be altered more readily than the deepest identity commitments yet more slowly than momentary preferences or moods. Someone might gradually revise their sense of being a “career-focused person” over a few years while maintaining more stable, slower priors about their fundamental values. At the same time, within a particular role, faster dynamics can fluctuate considerably—one can feel competent at work one day and overwhelmed the next—without immediately challenging the longer-range role prior. This layered flexibility allows the generative model to remain adaptive across diverse environments while preserving a core of temporal coherence that underwrites the sense of being a single, enduring self across many roles and phases of life.
Crucially, hierarchical time scales in self-representation are not just descriptive; they serve computational efficiency. By delegating fast-changing details to lower levels and reserving slower levels for abstract, time-averaged regularities, the bayesian brain reduces the complexity of inference. It does not need to reconsider fundamental identity questions every time there is a minor fluctuation in mood or context. Instead, it relies on stable high-level priors to interpret local variability as noise, routine, or situation-specific nuance. Only when discrepancies cross certain thresholds of magnitude, duration, or contextual generality are they escalated upward to prompt re-estimation of more abstract self-beliefs. This selective updating economizes computational resources while enabling the system to remain sensitive to meaningful changes in the environment and in its own capacities.
At the same time, this architecture creates characteristic vulnerabilities. If high-level priors are assigned excessively high precision, they may dominate lower-level evidence and enforce a rigid self-concept that is resistant to learning, even when confronted with clear, repeated counterexamples. If high-level priors are too imprecise or weakly integrated, lower-level fluctuations may gain undue influence, yielding a volatile sense of self that shifts dramatically with each new context or emotional state. The balance of precision across temporal scales thus shapes whether the self-model is experienced as stable yet responsive or as brittle and inflexible on the one hand, or as unstable and easily perturbed on the other. The phenomenology of identity—its solidity, its fragility, its openness to revision—emerges from how the generative model allocates confidence to different layers of its temporally structured expectations.
Learning, memory, and counterfactual self-generation
Learning within this framework is the gradual reshaping of the generative model’s temporally extended expectations about the self in light of prediction errors. Each encounter with the world provides evidence not only about external states but about how “someone like me” tends to perceive, feel, and act over time. When predictions about bodily states, social feedback, or emotional reactions are repeatedly violated, synaptic and systems-level plasticity adjust the relevant priors. Over many such encounters, the bayesian brain acquires structured regularities about its own behavior: what sorts of challenges it typically overcomes or avoids, which contexts it seeks or shuns, and how its affective states evolve in different scenarios. Learning thus amounts to a progressive refinement of the temporal profile of the self, compressing complex histories of interaction into compact expectations that can be deployed prospectively.
Memory, on this view, is not a passive storage of snapshots but an active, reconstructive process guided by these learned priors. When a past event is recalled, the generative model partially “re-runs” the inference that would have explained the sensory and internal states at that earlier time, given current beliefs about the trajectory of the self. The remembered scene is filled in by top-down predictions about what must have been the case, constrained by fragmentary traces and the present configuration of identity. Because these priors have been updated in the interim—shaped by later experiences and revised expectations about the future—recollection is inherently plastic. The same episode can be re-encoded or re-experienced in different ways depending on how it fits into the now-updated narrative of who one takes oneself to be, illustrating how learning reshapes memory by altering the assumed context in which past states are interpreted.
This reconstructive character of memory directly supports temporal coherence. Rather than treating each episode as an isolated record, the generative model reinterprets earlier events to minimize global prediction error across the entire life trajectory. Episodes that were once ambiguous or affectively neutral may be recast as decisive turning points if later developments make them predictive of current identity; others may fade into the background if they contribute little to the dominant self-hypotheses. Memory thus operates like a dynamic database undergoing continual schema revision: data are reorganized, compressed, or discarded to maintain a manageable, self-consistent structure that can efficiently inform future predictions. Forgetting, in this light, is not merely a failure of storage but often a byproduct of selective down-weighting of information that no longer serves the evolving generative model of the self.
The interplay between learning and memory is governed by precision-weighting. Experiences that generate highly precise, high-impact prediction errors—events that are surprising, emotionally intense, or recurrent across contexts—are more likely to drive updates at higher levels and be encoded as salient memories. Their influence extends beyond the immediate episode, reshaping priors about enduring traits, vulnerabilities, or capacities. Conversely, events that are consistent with existing high-level expectations may be encoded in a more schematic, low-precision manner, contributing marginally to already established patterns. Over time, this selective updating fosters a bias: the organism becomes especially sensitive to evidence that confirms or disconfirms its most consequential self-beliefs, and the memory system preferentially preserves episodes that bear on these beliefs, making the generative model increasingly tuned to identity-relevant contingencies.
Crucially, the same machinery that reconstructs memories from priors also generates counterfactual scenarios that extend beyond the actual past. Counterfactual self-generation involves running the generative model under altered constraints, asking in effect: “What would I be like if certain parameters of my life or character were different?” Inference proceeds by holding some aspects of the model fixed—such as core values or bodily structure—while varying others, like social role, skill set, or typical responses to stress. The model then simulates plausible trajectories emanating from these altered initial conditions, computing expected sensory, affective, and social consequences. These simulated trajectories are not mere fantasies; they are structured by the same learned dynamics that govern real behavior, constrained by the organism’s understanding of the world and its own capacities.
Counterfactual self-generation is supported by the temporal depth of predictive processing. Because the generative model already encodes transition probabilities between states—how one tends to move from confidence to anxiety in a certain context, or from curiosity to boredom over time—it can flexibly recompose these transitions into alternative sequences. For example, contemplating a future in which one persistently confronts social fears rather than avoiding them engages the model’s learned mappings between exposure, affective trajectories, and social outcomes, but organized into a different policy than the one currently enacted. The resulting imagined self is not arbitrary; it reflects a coherent reassembly of known dynamics under a new set of constraints on action and expectation. In this sense, counterfactuals exploit learned temporal regularities to explore the nearby space of possible selves.
These counterfactual trajectories feed back into both learning and memory. Imagined futures serve as “virtual evidence,” allowing the system to evaluate policies and self-trajectories without physically instantiating them. By mentally simulating alternative behaviors and outcomes, the generative model can assign expected prediction error and value to each path, biasing future exploration. If a counterfactual trajectory is assessed as likely to reduce long-term surprise—by enhancing predictability of social environments, bodily states, or internal conflicts—it may become a candidate for actual policy change. Subsequent real-world actions that partially instantiate this trajectory then provide concrete data, which either reinforce or disconfirm the simulated expectations, leading to further learning. In this way, counterfactual self-generation acts as a bridge between abstract possibility and embodied experimentation.
Counterfactual futures also reorganize autobiographical memory by altering which past events are treated as relevant. When the model considers a new possible self-trajectory—say, becoming a more assertive person—it searches the existing memory store for episodes that could support or contradict this imagined path. Past moments of assertiveness gain increased precision and salience, rising to consciousness more readily, while episodes of passivity may be down-weighted or reinterpreted as context-specific. The narrative of “who I have been” shifts to accommodate the imagined “who I could become,” yielding a partial retrofitting of memory around the counterfactual trajectory. Over repeated cycles, some counterfactual selves become increasingly anchored in recollection, acquiring the weight of plausibility through their integration into a reorganized autobiographical structure.
This bidirectional influence between memory and counterfactuals introduces a form of epistemic retrocausality without violating physical causation. Future-oriented expectations, even when purely simulated, impose new constraints on how the past is probabilistically reconstructed. If a newly entertained identity—such as “someone capable of recovery” after illness or trauma—proves successful at reducing prediction error in simulations and early enactments, it will shape which earlier experiences are interpreted as signs of resilience or latent strength. Those experiences, in turn, strengthen the prior for this emerging identity, making its associated future trajectories more credible and attractive. Learning thus propagates along loops that pass through both actual and counterfactual time, with the generative model continually adjusting its parameters to keep past memories, present states, and prospective selves in a workable equilibrium.
From the perspective of consciousness, the felt sense of “I could be otherwise” reflects the activation of these counterfactual generative processes. When one imagines different ways of acting in a current situation, or different life paths branching from a decision point, what is experienced is not a detached movie but a simulated inhabiting of alternate self-trajectories. Bodily, affective, and evaluative predictions are partially engaged: heart rate may shift, subtle muscular tensions arise, emotions of hope, dread, or relief flicker in anticipation of hypothetical outcomes. These phenomenological traces are the experiential counterparts of the generative model running alternative policies and updating its estimates of their costs and benefits. The space of possible selves is thus not an abstract catalog but a lived, affectively textured landscape generated in real time by predictive processing.
Because learning mechanisms operate on both actual and simulated experiences, the boundaries between “real” and “imagined” can blur at the level of the self-model. Repeatedly rehearsing a particular counterfactual—mentally practicing confident social interactions, envisioning oneself as a caregiver, or dwelling on catastrophic failures—can strengthen the associated priors, even if the corresponding behaviors are rarely enacted. The model becomes more ready to predict and explain the world through the lens of that imagined identity, altering attention, interpretation, and emotional response in situations that could confirm or disconfirm it. In turn, these biased interactions with the environment generate data that tend to validate the now-strengthened expectations, closing a feedback loop between imaginative rehearsal, selective sampling of evidence, and consolidation of new self-beliefs.
At the same time, the internal competition among counterfactual selves imposes constraints on learning. The generative model cannot fully endorse incompatible high-level trajectories without incurring chronic prediction error. As alternative futures are simulated, those that consistently generate high expected surprise—because they conflict with deeply entrenched values, bodily limitations, or social realities—are typically pruned or relegated to low-precision status. Others, which fit more harmoniously with existing structures while promising improved long-term predictability or value, gain precision and gradually displace older expectations. The resulting pattern of learning is therefore shaped by an internal selection process: counterfactual selves vie for dominance in the model, with memory and attention being allocated preferentially to those that best resolve the organism’s predictive and affective constraints across time.
This perspective recasts deliberate self-change—through therapy, education, or personal reflection—as an intentional manipulation of learning, memory, and counterfactual generation. Practices such as narrative reconstruction, cognitive reappraisal, or goal-setting introduce alternative self-trajectories into the generative model and supply structured contexts in which they can be simulated and tentatively enacted. By repeatedly revisiting past events from new interpretive standpoints and rehearsing new futures, individuals can shift the precision balance among competing priors, gradually reconfiguring which memories are salient, which counterfactuals feel realistic, and which patterns of action become default. Temporal selfhood, in this light, is not a fixed product but an ongoing outcome of learning processes that continuously weave together remembered histories and imagined futures into a dynamically stable yet revisable identity.
Implications for agency, identity, and psychopathology
Within this framework, agency is not an extra ingredient added on top of perception, but the way the generative model actively selects and enacts self-consistent trajectories through time. To experience oneself as an agent is to inhabit a configuration of temporally deep priors that specify what “I” typically do, can do, and ought to do in particular contexts. Actions are chosen as those policies that minimize expected prediction error across near and distant horizons, given these priors. When the bayesian brain’s predictions about intended movements, their sensory consequences, and their downstream social and affective outcomes are tightly aligned, actions feel owned, deliberate, and efficacious. The felt sense of “I did that” is thus the experiential counterpart of successful policy inference under a temporally extended self-model.
Disturbances of agency arise when there is systematic mismatch among levels of prediction or misallocation of precision. If high-level priors about what kind of person one is prescribe a certain course of action—being honest, caring, or brave—while lower-level predictions about bodily responses or social feedback anticipate failure, rejection, or danger, the generative model faces conflicting imperatives. It may oscillate between policies, produce ambivalent or half-hearted actions, or attempt to explain away discrepancies through post hoc narratives. Phenomenologically, this can manifest as indecision, akrasia (acting against one’s better judgment), or a sense of being “on autopilot.” In such cases, the experience of diminished agency reflects the model’s difficulty in finding a single low-error trajectory that satisfies incompatible temporal constraints on the self.
The allocation of precision across time scales plays a central role in shaping agency. When forward-looking priors about one’s capacities and typical outcomes are granted high precision, they strongly constrain which actions are even considered. A person with an entrenched expectation of social incompetence, for example, may never seriously entertain policies involving assertive engagement, because the generative model assigns high expected prediction error to such trajectories. As a result, exploratory actions that might falsify this self-belief are rarely sampled. Agency appears restricted not because options are physically unavailable, but because the predictive processing system has pruned them as implausible or too costly, narrowing the space of perceived possibilities and thereby the lived sense of what one “can” do.
This same machinery underwrites the normative and evaluative aspects of agency. The generative model does not predict only neutral states; it encodes expected value gradients—patterns of anticipated harm, benefit, shame, pride, and meaning—that contour the field of possible actions. Policies are not merely ranked by sensory predictability, but by their anticipated affective and social implications. Over time, these value-weighted priors crystallize as what one experiences as character, conscience, or practical wisdom. Acting “in line with my values” corresponds to selecting policies that both satisfy long-term predictions about being a certain kind of person and minimize anticipated affective surprise. When these long-range evaluative structures conflict—when, for instance, professional success and familial loyalty imply incompatible policies—the resulting dissonance in prediction and value may show up phenomenologically as guilt, ambivalence, or the sense of being pulled in different directions by different “parts” of oneself.
Identity, on this view, is the relatively slow-changing, high-level organization of priors that tie together diverse episodes into a unified trajectory. These priors encode enduring roles, traits, life projects, and value commitments, functioning as constraints on how new experiences are interpreted and integrated. To say “this is who I am” is to endorse a particular region of the generative model’s parameter space as the default explanation for past, present, and anticipated states. Because these high-level priors are temporally deep, they smooth over local variability, allowing momentary deviations to be absorbed as noise or context-specific anomalies. Identity stability is therefore not the absence of change, but the robustness of an organizing pattern that can accommodate perturbations without wholesale reconfiguration.
The same mechanisms that stabilize identity can, when biased, give rise to identity rigidity. If high-level priors about the self are granted excessive precision relative to incoming evidence, they may dominate inference to the point where contradictory data are persistently down-weighted or reinterpreted. A person might maintain a self-concept as “unlovable,” “always the strong one,” or “hopelessly flawed” despite sustained, clear counterevidence, because the cost of revising these priors is computed as too high in terms of expected global reorganization. This rigidity constrains counterfactual self-generation: alternative identities that conflict with entrenched priors are treated as unrealistic or even threatening, and thus are seldom explored in imagination or enacted in behavior. The result is an identity that feels solid but imprisoning, with agency experienced as constrained by a seemingly immutable “true nature.”
Conversely, when high-level identity priors are weak, inconsistent, or poorly integrated, the self can be experienced as unstable or diffuse. In such cases, lower-level fluctuations in mood, context, and social feedback exert disproportionate influence on self-ascription. A critical comment, a transient failure, or a brief emotional storm can trigger large swings in self-evaluation and behavior, because there is no dominant higher-order structure to buffer and contextualize these events. The generative model repeatedly reconfigures its long-range expectations in response to local surprises, producing a sense of being “different people” in different settings or of lacking any enduring core. Agency here may feel scattered or reactive: actions seem driven more by immediate pressures than by coherent, long-term projects that one recognizes as one’s own.
These considerations extend naturally into psychopathology, which can be reframed as patterns of chronic mismatch, mis-weighted precision, or maladaptive entanglement across temporal levels of self-representation. In depressive states, for example, high-level priors about expected future outcomes and one’s own efficacy may become globally pessimistic and overly precise: the bayesian brain predicts failure, rejection, or futility across many domains, and these predictions are treated as highly reliable. This temporal deepening of negative expectations reshapes memory, highlighting failures and downplaying successes, and constricts counterfactual futures to those that confirm hopelessness. Agency feels diminished not merely because energy is low, but because the generative model no longer registers many policies as viable paths to reduced long-term prediction error. The world appears as a landscape of foreclosed possibilities, and actions, when they occur, can feel pointless or externally compelled.
Anxiety disorders can be understood in terms of excessive precision on priors concerning threat and uncertainty at intermediate and long time scales. The generative model overpredicts the likelihood or severity of negative outcomes, especially those that are temporally extended—social humiliation, illness progression, financial ruin—and assigns high error costs to them. As a result, it preferentially simulates catastrophic futures and devotes significant resources to avoidance and hypervigilance policies aimed at preempting them. These policies, in turn, distort sampling of the environment, preventing disconfirmation of exaggerated threat expectations. Subjectively, agency is experienced as constrained by the need to avert imagined disasters, and identity can become organized around being “the anxious one,” “the responsible one,” or “the worrier,” roles that rationalize and stabilize the anxious self-model even as they maintain chronic distress.
Obsessive-compulsive phenomena illustrate how maladaptive loops can form between high-precision evaluative priors and lower-level action policies. Rigid expectations about moral purity, responsibility, or catastrophic harm—often encoded as temporally deep obligations extending indefinitely into the future—generate persistent prediction errors whenever the model cannot attain certainty that these obligations are met. Compulsions then emerge as policies designed to reduce immediate error (checking, cleaning, repeating), but they fail to resolve uncertainty at the relevant time scale: the long-term impossibility of guaranteeing absolute safety. Because each temporary reduction in anxiety is interpreted as evidence that the ritual is necessary, the generative model learns to treat compulsive policies as essential components of the self’s trajectory. Agency becomes entangled with these rituals, and actions that feel voluntary at the motor level are experienced as non-negotiable at the level of identity and obligation.
Dissociative experiences can be framed as breakdowns in the integration of temporal self-models across hierarchical levels. When events, affects, or actions generate prediction errors that cannot be reconciled with high-level priors—especially under conditions of trauma or overwhelming affect—the generative model may quarantine incompatible trajectories rather than update its core identity. This functional partitioning can lead to partially segregated sets of priors governing different contexts or states, with limited cross-talk between them. From the inside, this may be experienced as shifts into distinct “modes” of self, episodes of depersonalization, or a sense of watching one’s actions from the outside. Agency becomes fragmented: policies enacted under one configuration of the model may later be disowned or felt as alien when evaluated from another configuration that lacks access to the same predictive context.
Psychotic symptoms, such as delusions of control or thought insertion, can also be analyzed in predictive processing terms. Here, precision imbalances between predicted and actual sensory or cognitive states may lead the generative model to misattribute the source of actions or thoughts. If lower-level motor or cognitive processes generate outputs that are only weakly predicted by higher-level priors, the resulting large prediction errors must be explained. Rather than revising core identity priors, the system may posit external agents or forces as causes of these unexpected states (“someone else is moving my body,” “these thoughts are not mine”). The feeling that one’s agency has been usurped reflects the generative model’s attempt to restore coherence by relocating ownership of surprising trajectories to an external locus, maintaining a semblance of temporal order at the expense of accurate self-ascription.
Borderline patterns of instability in relationships, self-image, and affects can be interpreted as a chronic failure to stabilize high-level identity priors in the face of intense, rapidly fluctuating affective prediction errors. Interpersonal events are often assigned extremely high precision, causing each relational success or rupture to exert disproportionate influence on the global self-model. The result is rapid reconfiguration of long-term expectations: from idealization to devaluation of others, from empowerment to worthlessness in oneself. Counterfactual futures are repeatedly constructed and abandoned—fantasies of perfect union, total rejection, redemption, or annihilation—without sufficient time for any single trajectory to be consolidated through learning. Agency feels volatile and often reactive, as actions are driven by attempts to manage acute affective states rather than by stable, temporally extended projects that can anchor identity.
These psychopathological patterns highlight how vulnerabilities in temporal self-modeling are closely tied to the phenomenology of agency and identity. When long-range priors are overly rigid, the self becomes trapped in narrow, often painful trajectories that feel inevitable. When they are too weak or fragmented, the self lacks continuity and direction, leaving agency feeling arbitrary or externally dictated. In both extremes, consciousness of oneself as an agent among other agents is compromised: decisions feel either predetermined by an inflexible script or unmoored from any enduring standpoint. Interventions that restore a workable balance of precision across levels—loosening overly precise negative priors, strengthening incoherent or under-specified identity structures, and cultivating more flexible counterfactual self-generation—can be viewed as attempts to re-establish a viable space of self-consistent, low-error trajectories.
Therapeutic practices can thus be reinterpreted as targeted manipulations of the generative model’s temporal structure. Narrative therapies encourage re-authoring of one’s life story, explicitly inviting new ways of linking past, present, and future that challenge entrenched priors and open alternative paths. Cognitive-behavioral approaches systematically expose the system to prediction errors (for example, via behavioral experiments or exposure tasks) that are designed to weaken maladaptive high-precision beliefs and permit more adaptive priors to gain ground. Mindfulness-based interventions shift precision away from evaluative, long-range narratives toward immediate sensory and affective states, temporarily relaxing rigid identity constraints and allowing novel patterns of inference to emerge. Across these methods, the common thread is the deliberate reshaping of how the bayesian brain entangles its temporal hypotheses about the self, with the aim of expanding agency and stabilizing identity in ways that reduce suffering.
Importantly, these ideas blur the line between pathological and ordinary selfhood. Even in non-clinical populations, everyday struggles with procrastination, self-criticism, or role conflict can be understood as micro-failures of temporal coordination in the generative model. A person may hold partially incompatible long-term priors—ambitions pulling in one direction, relational loyalties in another—while lower-level policies attempt to satisfy both, producing cycles of avoidance, guilt, and reactive effort. Similarly, normative developmental transitions, such as adolescence or major life changes, involve large-scale updating of temporally deep priors about who one is and where one’s life is headed. During these periods, consciousness of agency and identity often feels labile or uncertain, reflecting the generative model’s active search for new, lower-error configurations of self across time.
