Personal World Models: Concepts and Applications

• Personal world models are persistent, updateable representations of an individual’s physical, digital, social, mental, and autobiographical environment, supporting personalized inference, simulation, planning, and action.
• Effective architectures combine multimodal perception, structured and episodic memory, action-conditioned prediction, perspective-aware social reasoning, provenance tracking, and explicit safeguards for uncertainty and privacy.
• Practical systems can improve personal-data retrieval, health and behavior analysis, assistive planning, human–robot interaction, and counterfactual decision-making, but must distinguish observed facts from beliefs, inferences, and hypothetical outcomes.

Personal world models are agent-specific, persistent, and updateable representations of a person’s physical, digital, social, mental, and autobiographical environment. They model not only external entities and events, but also the individual’s beliefs, goals, preferences, capabilities, memories, relationships, bodily condition, perceptions, and possible responses to alternative actions. The concept spans several research traditions: autobiographical scene-based reasoning, action-conditioned dynamics models, social and mental-state modeling, multimodal personal-data interpretation, structured memory systems, and causal representations of admissible counterfactual worlds. A personal world model is therefore more than a user profile, retrieval system, predictive model, or language-model context window: it is an evolving, perspectival model used for inference, simulation, planning, communication, and action.

1. Conceptual foundations and scope

A world model is commonly understood as an internal representation that allows an agent to predict how states may evolve under alternative actions. In its minimal form, it can be expressed as an action-conditioned transition distribution:

$s' \sim p(s' \mid s,a),$

where $s$ is a current state, $a$ an action, and $s'$ a possible successor state. A personal world model extends this structure by making the modeled state specific to an individual. It includes the person’s observations, uncertain beliefs, goals, preferences, bodily and affective conditions, memories, social relations, behavioral dispositions, and capabilities.

“Personal” has several distinct meanings in the literature. It may refer to a model centered on one agent’s egocentric observations and action consequences, as in humanoid action-conditioned video prediction (Ali et al., 1 Jun 2025). It may denote an autobiographical model whose reasoning emerges from an individual’s recorded experiences (Boloni, 2012). It may concern a user’s longitudinal physiological and behavioral data (Cosentino et al., 2024), or the representation of private beliefs, public expressions, and beliefs about others’ beliefs (Zimmaro et al., 20 Feb 2025). It may also refer to a target-specific, perspective-conditioned model of what a person can perceive and believe in a social environment (Alcedo et al., 26 Mar 2025, Fei et al., 29 Jul 2026).

A stronger definition treats a personal world model as a user-specific belief state:

$$b_t = p(s_t,z_t,m_t \mid o_{\leq t},a_{<t},g,\theta),$$

where $s_t$ is an external or shared state, $z_t$ includes bodily, affective, cognitive, and social variables, $m_t$ is personal memory, $o_{\leq t}$ are observations, $a_{<t}$ are prior actions, $s$0 represents current goals, and $s$1 contains person-specific traits, skills, preferences, and capabilities. This formulation is an extension of the world-model literature rather than a single formal definition adopted by all cited systems.

A personal world model should preserve the distinction between objective states, personal beliefs, and beliefs about beliefs. The statements “the appointment is tomorrow,” “the user believes the appointment is tomorrow,” and “the assistant believes that the user believes the appointment is tomorrow” describe different representational objects. Confusing them can produce incorrect personalization, unsafe recommendations, and inappropriate actions.

The intended scope is consequently broader than physical prediction. The relevant environment may include:

• Physical surroundings: bodies, rooms, objects, locations, motion, geometry, and affordances.
• Digital surroundings: files, applications, devices, calendars, messages, accounts, and workflows.
• Social surroundings: family, colleagues, institutions, norms, obligations, trust, and reputation.
• Mental state: knowledge, ignorance, beliefs, goals, intentions, emotions, values, expectations, and attention.
• Autobiographical continuity: remembered events, routines, preferences, relationships, and prior outcomes.

“Personal world model” is therefore best treated as an architectural category rather than a single algorithm. Current systems generally implement only subsets of this space.

2. Representational architectures

Autobiographical scenes and episodic alignment

Xapagy represents autobiographical experience as raw events organized into locally coherent scenes rather than as a single continuously updated model of reality (Boloni, 2012). Each scene contains story instances and verb instances. An instance belongs permanently to one scene; temporal continuity is represented through scene succession and explicit identity relations rather than by mutating one persistent object through time.

The system’s focus acts as working memory. Each focus element has a corresponding shadow: a weighted set of autobiographical instances or verb instances that resemble or structurally correspond to the current element. Shadowing is not direct recall. Past events cannot simply be reloaded into the focus; they influence current cognition through weighted alignments.

Headless shadows aggregate autobiographical evidence into structured expectations or inferences. They can represent a likely successor event, a hidden predecessor, an inferred relation, an elaboration, a summary, or a continuation. Their support is episodic and relational rather than derived from explicitly extracted general rules. A wolf eating an animal, a tiger eating prey, and a wolf eating a girl can support structurally related predictions even when the surface labels differ.

This architecture provides a distinctive model of personal continuity. Identity is relational and fragmented: an individual at different times or in different episodes is represented by multiple scene-local instances connected by identity relations. Recall creates a new story rather than replaying an old one. The resulting self is not a single immutable entity but a network of situated representations.

Shared dynamics and latent state

Meta-world models use environment-specific visual encoders and a shared recurrent dynamics model (Wu et al., 2018). For environment $s$2, the model is:

$s$3

where $s$4 maps observations into a latent representation and $s$5 predicts action-conditioned latent dynamics shared across environments. The visual encoders and decoders differ because the observation spaces differ, while the recurrent model captures common physical structure.

The architecture uses a VAE-like visual model with a latent dimension of $s$6 and an LSTM memory model with $s$7 hidden units. The LSTM receives the current latent representation, its recurrent state, and the agent’s action. It predicts a diagonal Gaussian distribution over the next latent state. Reconstruction and prediction losses are optimized alternately, with a distribution-matching regularizer encouraging compatible latent spaces across environments.

This model contributes an important component of personal world modeling: a compact predictive representation of action-conditioned environmental dynamics. Its mirror-environment experiment supports a limited form of action ownership or self/other discrimination. It does not, however, model biography, persistent identity, preferences, body morphology, autobiographical memory, or individual adaptation.

Hierarchical multimodal and generative representations

A broader world-model specification advocates combining continuous and discrete representations, multiple modalities, hierarchical prediction, and action-conditioned generation (Xing et al., 7 Jul 2025). Continuous representations can encode sensory and bodily variation, while discrete tokens can support compositional reasoning over objects, events, concepts, norms, goals, and memories. A hierarchical system may contain low-level perceptual dynamics, intermediate object and event representations, and higher-level planning or counterfactual reasoning.

Humanoid World Models implement a lower-level version of this principle as short-horizon egocentric video prediction conditioned on robot actions (Ali et al., 1 Jun 2025). Masked latent-video models and flow-matching models generate future observations from past frames and proposed action sequences. Their demonstrated scope is short-term visual dynamics, not a persistent model of personal identity, social context, goals, or memory.

The distinction between visual prediction and world modeling is emphasized by critiques of world models: a video generator or predictive embedding becomes a world model only when it supports actionable simulation, persistent state, controllability, and planning (Xing et al., 7 Jul 2025). Image quality alone does not establish causal or physical adequacy.

Structured state and executable rules

Web World Models separate code-defined state and rules from model-generated narrative or semantic content (Feng et al., 29 Dec 2025). Their state is represented as:

$s$8

where $s$9 is the deterministic, code-defined state and $a$0 is model-generated perceptual or narrative state. Code applies the action transition:

$a$1

after which a model generates contextual content conditioned on the updated state.

Typed interfaces, JSON schemas, deterministic transitions, validation, caching, and fallback templates provide persistence and controllability. Applied to personal systems, this suggests separating confirmed preferences, calendar commitments, permissions, task states, and resource constraints from model-generated hypotheses, summaries, explanations, and plans. The generated layer should not silently modify protected personal state.

3. Personal memory, identity, and evidence

Personal world models require memory systems that distinguish observations, episodes, semantic summaries, procedures, preferences, relationships, and hypotheses. A statement explicitly provided by a person differs from a preference inferred from repeated behavior, which differs from a model-generated interpretation. Each memory should ideally retain provenance, confidence, timestamp, sensitivity, validity interval, and whether it has been confirmed.

Digital personal environments

HippoCamp evaluates agents in user-centric file systems containing more than 2,000 heterogeneous files and 42.4 GB across text, documents, images, video, and audio (Yang et al., 1 Apr 2026). Its tasks require retrieval, evidence localization, temporal reasoning, entity binding, cross-modal integration, and user profiling. Profiling includes preferences, behavioral patterns, scheduling policies, retrospective events, and workflows.

The benchmark demonstrates that personal information is distributed across files, modalities, dates, people, and workflows. A personal model must therefore represent not merely that a fact occurs in a file, but who it concerns, when it holds, how it relates to other evidence, and whether multiple weak observations support a stable inference.

HippoCamp distinguishes factual retention from profiling. Agents can retrieve relevant documents yet fail to construct a correct user-level model. Common failures include retrieval mismatch, evidence avoidance, hallucinated paths, entity misbinding, and verification deficits. Profiling is substantially harder than factual retention, indicating that personal modeling requires abstraction over longitudinal and relational evidence rather than retrieval alone.

Visual personal context

Personal Visual Context Learning studies whether frozen multimodal models can use private visual evidence supplied at inference time (Xue et al., 11 May 2026). Personal-VCL-Bench evaluates person identity, object ownership, procedural behavior, personalized deviation detection, social relations, and egocentric wearer identification across 2,255 context-query instances.

The Agentic Context Bank organizes visual evidence into typed entries for appearance, owned objects, and behavior. Each entry links a natural-language descriptor to supporting frames or temporal video spans. New evidence can add, confirm, revise, or retract an entry. At query time, the model first uses a textual view of the bank and selectively requests raw visual evidence for verification.

This architecture provides a lightweight form of memory consolidation and query-adaptive grounding. Its limitations are equally important: it lacks a formal entity graph, calibrated uncertainty, explicit temporal validity, spatial maps, social memory, predictive dynamics, and lifelong operation. It is a memory and evidence-grounding layer rather than a complete personal world model.

Autobiographical identity

The Xapagy architecture treats autobiographical memory as the foundation of cognition rather than as a subsystem alongside semantic and procedural memory (Boloni, 2012). Concepts acquire meaning through their occurrence in concrete events, and identity is represented through explicit relations between scene-local instances. This permits multiple incompatible narratives and perspectives to coexist without global consistency resolution.

A personal system adopting this approach would represent identity as a temporally and contextually structured network. The same person may have multiple representations corresponding to different periods, roles, bodily conditions, or narrative contexts. Such fragmentation can preserve perspective and change, but it requires explicit identity resolution and temporal validity to prevent accidental conflation.

Health and physiological state

PH-LLM interprets longitudinal wearable data for sleep and fitness (Cosentino et al., 2024). It combines tables, graphs, text summaries, demographics, and aggregated statistics, including sleep, activity, heart rate, HRV, respiratory rate, training load, subjective readiness, and muscle soreness. The model produces personalized insights, etiological hypotheses, recommendations, and predictions of self-reported sleep outcomes.

The system demonstrates multimodal personal-state inference, but its personalization is contextual rather than enduring. It receives a bounded window of data and generates an answer; it does not maintain persistent memory, a recurrent personal state, a validated causal model, or long-term adaptation. The PRO experiment reduces 15 days of sensor data to temporal means and variances, preserving limited summary statistics but discarding much of the ordering and transition structure.

4. Perspective, social cognition, and mental state

A physical model alone cannot reliably predict human behavior because people act on the basis of what they know, believe, want, intend, feel, and consider permissible. Mental World Modeling formalizes a coupled physical-mental state (Fei et al., 29 Jul 2026):

$a$2

The physical component represents objects, characters, spatial relations, motion, visibility, environmental conditions, and perceptible signals. The mental-social component represents individual beliefs, attention, goals, intentions, emotions, dispositions, norms, constraints, identities, relationships, and social atmosphere.

A target-specific observation function renders only what a particular agent can perceive and infer. It incorporates perceptual access, social role, prior knowledge, relationships, and Theory-of-Mind capacity. Facts inaccessible to the target are represented as unknown rather than supplied from the global state. This prevents perspective leakage: giving an agent information that the modeled person could not have observed.

Mental World Modeling separates the global simulator from each agent’s subjective model. The global state may contain the true physical state and all agents’ mental states, while the target observation contains that target’s partial and potentially false beliefs. Candidate actions are decomposed into physical carriers and mental or social meanings. The transition then updates physical and mental states jointly.

Mentis, its training-free and inspectable implementation, parses a structured state, generates a target-specific observation, decomposes candidate actions, simulates coupled successor states, and evaluates branches for mental consistency, physical plausibility, and social appropriateness (Fei et al., 29 Jul 2026). On Menti-Bench, full mental world modeling achieved an average F1 of 87.9, compared with 63.3 for direct answering and 80.3 for a free-text state representation. Removing mental information reduced average F1 by 12.1 points, removing physical information by 16.5 points, and decoupling transitions by 6.4 points.

This framework is related to, but distinct from, a strong personal world model. It represents target-specific perspectives within a scenario but generally lacks persistent identity, autobiographical memory, person-specific adaptation, long-term relationships, and longitudinal preference learning.

Personal, expressed, and social beliefs

The PES meta-model separates private beliefs, expressed beliefs, and beliefs about others’ beliefs (Zimmaro et al., 20 Feb 2025). For agent $a$3, personal belief is $a$4, expressed belief is $a$5, and the belief about another agent $a$6 is $a$7. The information pathway is:

$a$8

This distinction supports strategic communication, conformity, preference falsification, false consensus, false uniqueness, self-reinforcement, and inaccurate social perception. It also provides a minimal theory-of-mind architecture: an agent models not only its own position but another agent’s public behavior and its interpretation of that behavior.

PES generalizes Voter, Ising, DeGroot, bounded-confidence, and related opinion-dynamics models by allowing private, public, and social states to differ. It is not a complete personal world model because it omits rich perception, episodic memory, physical state, explicit goals, detailed communication channels, causal models, and action policies. It is best understood as a social-belief subsystem.

Perspective-shifted navigation

Perspective-Shifted Neuro-Symbolic World Models represent a robot’s own factored belief and an estimated belief of another agent (Alcedo et al., 26 Mar 2025). The robot’s belief $a$9 is transformed through a perspective-shift operator:

$s'$0

where $s'$1 is the other agent’s observed configuration and $s'$2 is the learned world model. The operator simulates what the other agent might observe from its location and sensing capabilities.

The architecture uses discrete latent belief factors, a learned forward model, an observation decoder, an inverse model, and a policy conditioned on both $s'$3 and $s'$4. It implements a limited form of Theory of Mind through mental simulation, but does not model identity-specific traits, persistent interpersonal memory, communication-based belief revision, or nested beliefs beyond what could be obtained by recursively applying the perspective operator.

5. Prediction, intervention, and counterfactual structure

A personal world model must predict not only likely next observations but also the consequences of alternative actions. It should distinguish ordinary uncertainty over the current state from uncertainty about how alternative worlds are coupled.

WorldKernel formalizes a world model as a coupling kernel over admissible possible worlds (Rovai, 9 Jun 2026):

$s'$5

where $s'$6 and $s'$7 are complete admissible worlds and $s'$8 represents evidence-conditioned uncertainty. The diagonal $s'$9 represents ordinary posterior plausibility. Off-diagonal structure represents relationships among alternatives, especially cross-world couplings relevant to counterfactual queries.

For binary potential outcomes $$b_t = p(s_t,z_t,m_t \mid o_{\leq t},a_{<t},g,\theta),$$0 and $$b_t = p(s_t,z_t,m_t \mid o_{\leq t},a_{<t},g,\theta),$$1, the marginal probabilities $$b_t = p(s_t,z_t,m_t \mid o_{\leq t},a_{<t},g,\theta),$$2 and $$b_t = p(s_t,z_t,m_t \mid o_{\leq t},a_{<t},g,\theta),$$3 may be identifiable while the joint probability $$b_t = p(s_t,z_t,m_t \mid o_{\leq t},a_{<t},g,\theta),$$4 remains unidentified. Two models can agree on all observed and interventional marginals while disagreeing about what would have happened to the same individual under an alternative intervention.

For personal reasoning, this distinction affects questions such as:

• What would the person have done if they had received different advice?
• Would the disagreement have occurred if the earlier message had not been sent?
• Would the person still have trusted a colleague under an alternative history?
• Which future trajectories remain jointly consistent with their values and relationships?

A strong predictor may return a single number for such a query even when the evidence supports only an interval. WorldKernel proposes positive-semidefinite constraints, ontology constraints, and targeted evidence acquisition to maintain admissible ranges rather than arbitrary point estimates. Its positive-semidefinite relaxation is polynomial in the number of variables, whereas the exact response-type representation can be exponential.

The framework also introduces “scars”: persistent constraints learned from encountered infeasibilities. Targeted scars identify the worst-case configuration for a query, locate the violated constraint, store it, and recompute the bound. In a personal system, analogous active acquisition could select which memory to retrieve, which ambiguity to clarify with the person, which preference to verify, or which intervention would most reduce uncertainty about a consequential outcome.

This counterfactual perspective complements predictive world models. Video prediction and recurrent dynamics models estimate likely future observations, while coupling-based models ask whether alternative futures are jointly consistent with shared causal, logical, physical, and personal constraints.

6. Engineering, evaluation, safety, and limitations

Hybrid architectures

Several works converge on hybrid designs combining learned perception and prediction with explicit symbolic, typed, or constraint-based layers. Web World Models place deterministic code beneath model-generated narratives (Feng et al., 29 Dec 2025). Perspective-shifted navigation uses discrete latent factors and a learned forward model (Alcedo et al., 26 Mar 2025). Critiques of world models advocate hierarchical mixed continuous/discrete representations and observation-grounded generative objectives (Xing et al., 7 Jul 2025). The reliability argument in “AI and World Models” calls for explicit physical, social, and mental models surrounding neural generators (Worden, 25 Jan 2026).

A practical personal world model could contain:

1. Multimodal interaction gateway: text, audio, images, video, sensors, applications, calendars, and user commands.
2. Grounded state estimator: entities, events, locations, actions, relationships, bodily state, and temporal intervals.
3. Memory system: episodic, semantic, procedural, relational, preference, and task memory.
4. Shared–private knowledge layer: general model knowledge separated from user-specific data.
5. Symbolic–latent reasoning core: explicit representations for verification and latent representations for continuous multimodal dynamics.
6. Predictive simulator: physical, digital, social, and mental counterfactuals.
7. Planner and action gateway: authorization, safety constraints, reversibility, and confirmation.
8. Provenance and governance layer: confidence, source, timestamp, expiry, deletion, correction, and audit.
9. Multimodal generator: explanations, timelines, spatial previews, alternatives, and clearly labeled hypothetical content.

Common ground and communication

A personal model used in dialogue must maintain an explicit or implicit model of common ground: the overlapping subset of the user’s and assistant’s physical, social, and mental world models (Worden, 25 Jan 2026). Common ground includes shared facts, salient entities, social roles, user knowledge, unresolved uncertainty, goals, expectations, and beliefs about the assistant.

The assistant should distinguish what the user knows from what the assistant knows, what the user believes from what is established, and what the user expects from what the system can actually do. Divergence should trigger conversational repair rather than silent continuation. This is particularly important because LLMs lack a stable persistent common ground and can generate plausible language without maintaining task state, evidence, or commitments.

Evaluation

No single metric captures personal world-model quality. Relevant evaluation dimensions include:

• Observation prediction: perceptual and temporal fidelity.
• State prediction: object, location, activity, bodily, social, and mental-state accuracy.
• Action consequences: success, failure, safety, and causal validity after interventions.
• Long-horizon consistency: error accumulation, memory persistence, and commitment tracking.
• Counterfactual validity: feasible identified sets and consistency across related queries.
• Evidence grounding: retrieval, localization, provenance, and verification.
• Entity continuity: binding observations to the correct person, object, event, or relationship.
• Personalization: improvement on held-out user-specific tasks.
• Uncertainty calibration: negative log likelihood, coverage, abstention, and risk-sensitive performance.
• Privacy and governance: leakage, consent, deletion, access control, and third-party protection.
• User correction: ability to inspect, revise, contest, and roll back model state.

HippoCamp shows that retrieval and answer accuracy can diverge sharply (Yang et al., 1 Apr 2026). Personal Visual Context Learning shows that supplying more visual context does not reliably improve performance and that structured, selective evidence access can outperform unstructured context concatenation (Xue et al., 11 May 2026). These results support evaluating intermediate artifacts and processes, not only final answers.

Safety and governance

A personal world model aggregates highly sensitive information about health, finances, location, relationships, routines, vulnerabilities, beliefs, and behavior. It can therefore amplify both assistance and manipulation. Safety requirements include local-first processing where possible, encryption, field-level access control, retention policies, sensitivity labels, user-visible provenance, selective deletion, per-tool permissions, and explicit confirmation for consequential actions.

The system should distinguish observations, confirmed facts, inferences, plans, and actions. A temporary conversational statement should not automatically become a permanent preference. Inferred mental states should remain contestable. Data concerning third parties should not be modeled merely because it is accessible through the user’s files or devices.

A reliable outer architecture should constrain neural components through typed interfaces, deterministic transitions where appropriate, schema validation, executable action boundaries, safety vetoes, and fallback behavior. Determinism is useful for reproducibility, but personal states must remain temporally versioned and revisable rather than frozen by a seed. Generated content should never be treated as an observation without explicit provenance.

Fundamental limitations

Current systems leave major aspects of personal world modeling unresolved:

• Persistence: most systems operate over bounded contexts or episodes rather than lifelong memory.
• Identity: people and objects may be misbound across files, scenes, or visual observations.
• Temporal reasoning: routines, change points, and long-term transitions are incompletely represented.
• Causal inference: predictive association is often mistaken for intervention-valid reasoning.
• Uncertainty: many systems generate point estimates or narratives despite unidentified alternatives.
• Social realism: norms, trust, relationships, and intentions are culturally and contextually variable.
• Preference drift: users change through health events, relationships, work, and environmental transitions.
• Privacy: detailed personal models create risks of surveillance, profiling, unauthorized inference, and data leakage.
• Evaluation: benchmarks are generally static, curated, narrow, or scenario-specific.
• Computational complexity: exact modeling of all personal histories, futures, and higher-order beliefs is generally intractable.

The principal research challenge is integrating persistence, perception, memory, dynamics, belief, social cognition, causal reasoning, and governance without collapsing the distinction between observed fact and inferred hypothesis. A complete personal world model would need to predict actionable possibilities while preserving the individual’s perspective, uncertainty, autonomy, privacy, and authority over the representation of their own life.