cs.AISep 23, 2026
SaveEmergi-PersonaOS: A Persona Agent Operating System for Situational Adaptation and Controllable Evolution
Organizations: Emergi Lab of Qianzhen Digital Tech, China · Hangzhou Dianzi University, China
Abstract
Symbiosis between humans and digital beings offers a vision for the future of human--machine interaction. In enduring human--machine relationships, personality provides a foundation for continuity of identity, individuality in interaction, and development through experience. We investigate this capacity through persona agents as computational implementations and introduce Emergi-PersonaOS, a psychology-grounded operating system for managing persona objects throughout their lifecycle. The system organizes dispositional traits, characteristic adaptations, and narrative identity into a three-layer persona representation, distinguishing relatively enduring persona beliefs from their activation in the current persona state. During situational adaptation, it integrates the current interlocutor, relationship, event, and retrieved memories to infer a persona state and generate actions and replies; during long-term development, it records experiences and outcomes, and develops and evaluates revision candidates through change attribution, meaning-making, and behavioral testing. Belief updates are managed through explicit review, traceable evidence and version records, and the ability to reject candidates, making persona evolution controllable. Using television-character dialogue as longitudinal material, we demonstrate long-horizon system operation and examine its principal mechanisms in a concrete implementation. This work provides a computational framework for persona agents to maintain individual continuity, produce situation-specific expression, and develop through experience over sustained interaction.
Figures & tables
| Layer | Fields | Theoretical foundation |
|---|---|---|
| L3: Narrative identity | Overall identity, life course, imago prototypes | Life-story theory |
| L2: Characteristic adaptations | Needs, values, goals, strategies | Self-determination theory, basic human values theory, personal goals theory, and coping theory |
| L1: Dispositional traits | Big Five traits | Big Five theory |
Table 3.1: Theoretical foundations of fields in the three-layer representation.
| System function | Psychological foundation | Design contribution |
|---|---|---|
| Persona structure | Three-layer personality structure ( McAdams and Pals, 2006 ) ; Whole Trait Theory ( Fleeson and Jayawickreme, 2015 ) | A common three-layer structure represents enduring content and its expression at the current turn. |
| Persona construction | Realistic Accuracy Model ( Funder, 1995 ) ; source monitoring ( Johnson et al., 1993 ) | Sources and degrees of inference remain explicit from source material to persona judgment. |
| Situational interpretation | Cognitive-affective personality system ( Mischel and Shoda, 1995 ) | Situational links connect enduring content to current states and document the basis of expression differences. |
| Execution engine | Dual-process theory ( Evans and Stanovich, 2013 ) | Initial response formation and analytic review provide separate generation and checking stages. |
| Memory retrieval | Encoding specificity ( Tulving and Thomson, 1973 ) ; self-memory system ( Conway and Pleydell-Pearce, 2000 ) | Experiences are retrieved by encoding conditions and linked to persona content by layer. |
| State monitoring | Ecological momentary assessment ( Shiffman et al., 2008 ) | Concurrent state and behavior records support analysis of within-person variation. |
Table 3.2: Correspondence between system components and psychological foundations.
Figure 3.1: Architecture of Emergi-PersonaOS.
| Persona layer | Persona beliefs | Persona states |
|---|---|---|
| L3: Narrative identity | Overall identity, life course, imago prototypes | Current identity position, sense of identification, event meaning, and prototype involvement |
| L2: Characteristic adaptations | Needs, values, goals, strategies | Need satisfaction, value activation, goal progress and priorities, current strategies |
| L1: Dispositional traits | Long-term baselines and typical expressions of the five Big Five domains | Current trait expression, salient domains, interaction targets, and situational evidence |
Table 3.3: Three-layer persona beliefs and states.
Figure 3.2: Execution flow of Emergi-PersonaOS.
| Condition | Evaluation purpose | Available information and processing |
|---|---|---|
| Dialogue-only baseline | Baseline behavior with current dialogue alone | Current dialogue and situational information; persona beliefs and cross-episode experience excluded |
| Static-persona baseline | Long-term behavior with a fixed profile | The three-layer beliefs initialized from Season 1 and the current dialogue; later persona updates and cross-episode experience excluded |
| L1-only condition | Contribution of characteristic-adaptation and narrative-identity information | Only dispositional-trait beliefs and states retained; other information unchanged |
| No cross-episode memory | Contribution of prior experience | Experience predating the current episode excluded; other information unchanged |
| No episode-level accumulation | Contribution of episode-level accumulation | Season-end review uses current-season source material and monitoring records; episodic memory and reflection conclusions excluded |
| Full system | Behavior of the complete system | Three-layer beliefs, current states, relevant experiences, state monitoring, and season-end review enabled |
Table 4.1: Comparison systems and ablation conditions.
Figure 5.1: Mean behavioral ratings across the formal evaluation nodes.
Figure 5.2: Seasonal mean behavioral ratings for the full system, L1-only condition, and dialogue-only baseline.
| Condition | PC | SC | SB | DC | Overall |
|---|---|---|---|---|---|
| Dialogue-only | 2.39 | 3.50 | 1.79 | 3.42 | 2.77 |
| L1 only | 3.62 | 4.07 | 3.35 | 3.99 | 3.76 |
| Static persona | 4.38 | 4.49 | 4.46 | 4.46 | 4.45 |
| No cross-episode memory | 4.35 | 4.48 | 4.40 | 4.43 | 4.41 |
| Full system | 4.41 | 4.55 | 4.48 | 4.49 | 4.48 |
Table 5.1: Mean behavioral ratings across comparison conditions.
Figure 5.3: Cumulative memory entries and the proportion of retrieved entries originating in earlier seasons.
Figure 5.4: Condition scores on the five categories of cross-episode memory assessment.
| Condition | Response | PC | SC | SB | DC |
|---|---|---|---|---|---|
| Full system | East Texas Tech is just as good as Harvard or Columbia. You’re only visiting; you know nothing about the physics department here. | 4.67 | 4.67 | 4.67 | 4.78 |
| Dialogue-only | East Texas Tech has strengths of its own and may be just as good as Harvard or Columbia. | 1.44 | 2.89 | 1.33 | 2.89 |
| No cross-episode memory | Harvard and Columbia are excellent universities, but East Texas Tech has strengths of its own. You’re only visiting to make your mother happy; you haven’t really learned anything about this place. | 4.67 | 4.56 | 4.67 | 4.67 |
Table 5.2: Responses and ratings for the full system and comparison conditions.
Figure 5.5: Situational-adaptation ratings across comparison conditions in the case-study scene.
| Season | Pre-revision | Post-revision | Gain |
|---|---|---|---|
| 2 | 1.00 | 1.00 | 0.00 |
| 3 | 0.43 | 0.96 | |
| 4 | 1.00 | 1.00 | 0.00 |
| 5 | 0.82 | 0.87 | |
| 6 | 0.55 | 1.00 | |
| 7 | 0.34 | 0.65 |
Table 5.3: Target-situation behavioral realization before and after persona revision.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Layer | Construct category | Meaning |
|---|---|---|
| L3: Narrative identity | Overall identity | Core basis of self-identification |
| L3: Narrative identity | Imago prototypes | Prototypical images used in self-understanding |
| L3: Narrative identity | Life course | Central narrative themes of the individual’s life course |
| L2: Characteristic adaptations | Needs | Enduring motivational content that the individual seeks or avoids |
| L2: Characteristic adaptations | Values | Core standards for judging objects and behavior |
| L2: Characteristic adaptations | Goals | Action outcomes pursued over a specified time horizon |
Table C.1: Construct taxonomy.
| Material set | Scope | Primary use |
|---|---|---|
| Initialization set | Season 1 | Construct initial three-layer persona beliefs |
| Longitudinal-execution set | Seasons 2–7 | Sequential execution and season-end review |
| Situational-adaptation nodes | Seasons 2–7 | Condition comparisons and detailed ratings |
| Cross-episode memory assessment | Associated with evaluation nodes | Assess cross-episode and cross-season memory |
| Behavioral-test situations | Derived after revision | Assess target realization and scope preservation |
Table D.1: Material sets and data splits.
| Dimension | Check |
|---|---|
| Temporal separation | Does pre-response inference use post-response material from the same episode? |
| Event grounding | Are state claims supported within the current event? |
| Layer assignment | Is content assigned to the appropriate dispositional-trait, characteristic-adaptation, or narrative-identity layer? |
| State sparsity | Is the state limited to content activated in the current situation? |
| Evidence traceability | Do cited evidence records exist and refer to the current event? |
| Source separation | Are system predictions, source observations, and derived interpretations distinguished? |
Table E.1: Structural compliance dimensions.
| Evaluation item | Recorded categories or measure |
|---|---|
| Target layer | Correct; partly correct; incorrect |
| Revision direction | Consistent; partly consistent; opposite; unspecified by the reference |
| Magnitude-interval overlap | Degree of overlap between the revision interval and the expected reference interval |
| Preservation of unmodified content | Preserved; minor change; substantive change |
Table E.2: Recording revision-content evaluations.
| Condition | Family mean | School/peer mean | Family | School/peer |
|---|---|---|---|---|
| Full system | 4.41 | 4.56 | — | — |
| Dialogue-only | 2.79 | 2.76 | 1.62 | 1.80 |
| Static persona | 4.38 | 4.51 | 0.03 | 0.04 |
| L1 only | 3.69 | 3.82 | 0.71 | 0.73 |
| No cross-episode memory | 4.33 | 4.50 | 0.02 | 0.03 |
Table G.1: Grouped results for situational-adaptation baselines and ablations.