Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Artificial Identity and Persona Continuity

Abstract. Artificial identity and persona continuity concern the persistence and revision of an identifiable artificial system's documentary, functional, narrative, and relational organization across sessions, model upgrades, and changing social circumstances. This fundamental research monograph establishes operational distinctions between persistent identifiers, historical records, persona specification, measurable behavior, and public recognition. It synthesizes the psychology of autobiographical memory, narrative identity and anthropomorphism with primary computational studies of personalized dialogue, multi-session memory, generative agents, persona consistency, role-play evaluation, and structured long-term retrieval. Experimental paradigms, benchmark families, validity threats, causal attribution, provenance, migration, branching, privacy, and adaptive correction receive critical examination. The synthesis explains why service-level identity, stored records, behavioral stability, and human recognition require separate evidence while remaining causally and socially connected. It concludes with an operational research program for longitudinally evaluating artificial personas amid system change. The method is a scholarly narrative synthesis; no original participant study, registered systematic review, meta-analysis, or clinical recommendation is claimed.

Keywords: artificial identity; persona continuity; artificial persona; artificial memory; autobiographical memory; persona consistency; long-term conversational memory; model migration; identity persistence; behavioral continuity; persona drift; agent identity; anthropomorphism; human–AI interaction; structured memory; artificial authorship; longitudinal evaluation

Artificial Identity as an Empirical Research Object

Artificial identity and persona continuity concern the persistence of an identifiable artificial system's publicly attributable organization across successive interactions, memory revisions, model substitutions, and changes in operational context. Identity in this inquiry is a family of testable relations among records, implementations, response dispositions, and recognized roles. A service can preserve an account identifier while replacing its generative model; a model can preserve characteristic outputs while losing its record of a particular interlocutor. These cases require separate observations. The target phenomenon is the temporally extended configuration of an artificial agent that can be tracked, evaluated, revised, and recognized under explicit continuity criteria Zhang et al. (2018) Park et al. (2023).

The scientific questions arise at several levels. What is held constant: a database key, a model checkpoint, a narrative specification, a collection of personal memories, a set of response policies, or an externally recognized persona? What counts as successful persistence when new knowledge changes legitimate behavior? What evidence distinguishes an agent that remembers prior exchanges from one that plausibly guesses them? Which transformations preserve a usable identity and which produce a substituted or forked agent? Long-term dialogue benchmarks and persona evaluation frameworks now permit portions of these questions to be addressed experimentally, while the philosophical and social dimensions require different methods Maharana et al. (2024) Samuel et al. (2025).

The method of this monograph is a critical narrative synthesis of psychology, human–computer interaction, computational agent research, and natural-language processing. Original work on autobiographical self-organization is used as a comparative source of constructs, while engineering studies provide tests of persistent conversational memory, character fidelity, and behavioral stability. The evidential standard is explicit: an observed performance score establishes performance under the study protocol; a continuity claim requires longitudinal comparison, controls for retrieval and prompts, and a declared identity criterion. The inquiry develops an integrative framework as a research interpretation, without claiming a new experimental discovery Conway & Pleydell-Pearce (2000) Frisch & Giulianelli (2024).

Four Analytically Distinct Forms of Continuity

Identifier continuity is the persistence of an administrative or cryptographic marker across operations. A stable account, deployment name, or agent handle makes records joinable and enables attribution, audit, and access control. This can survive wholesale replacement of weights, tools, memory, and instructions. Its evidential significance is therefore infrastructural: it shows that a provider treats events as belonging to one managed entity. It gives researchers a necessary linkage variable for longitudinal studies but does not itself demonstrate preservation of a characteristic conversational disposition or of the agent's accumulated knowledge Packer et al. (2023) Zhong et al. (2024).

Record continuity concerns the preservation and accessibility of relevant history. Stored transcripts, event graphs, user preferences, versioned persona facts, and retrieval indexes can support recall across sessions. Record continuity requires more than a stored file: information must remain semantically retrievable, temporally valid, correctly attributed, and appropriately used. A model with extensive archival storage can still fail a question about an earlier event because the wrong episode is retrieved or because it confuses a user's experience with its own attributed history. LoCoMo isolates several such long-range challenges and cautions against treating storage capacity as equivalent to recall competence Maharana et al. (2024) Park et al. (2023).

Persona continuity is the persistence of a recognizable configuration of norms, voice, role, factual commitments, and action preferences across circumstances. Behavioral continuity is the measured stability or principled evolution of responses under comparable prompts and social situations. These dimensions interact, yet they can diverge: an assigned character may sound consistent while contradicting its own past, and an accurate memory system may support an evolving role whose presentation changes appropriately. PersonaGym explicitly tests multiple kinds of persona adherence; interaction studies find variability in trait expression and linguistic accommodation. A scientifically useful identity account measures these dimensions separately before considering their joint effects Samuel et al. (2025) Frisch & Giulianelli (2024).

Historical Foundations in the Psychology of Self and Memory

The psychology of human identity developed through the study of memory, self-description, social recognition, narrative meaning, and individual development. This tradition supplies distinctions that are useful when studying constructed artificial personas. Conway and Pleydell-Pearce's self-memory system conceptualizes autobiographical remembering as a process shaped by self-relevant goals and organized knowledge, rather than passive access to a complete historical recording. Its contribution to artificial-identity research is conceptual and methodological: test the relation between ongoing priorities, retrieved evidence, and a coherent account of the past. Artificial retrieval systems must be evaluated on their own mechanisms Conway & Pleydell-Pearce (2000).

McAdams's actor–agent–author account explains human selfhood through social roles, motivated projects, and autobiographical narration across development. For artificial systems, actor, agent, and author can be reformulated as observable task roles, control policies, and generated historical accounts, provided that these are operational descriptions and the human developmental theory is not transferred wholesale. The distinction directs attention to how an artificial persona can maintain a public role even when underlying policies and memory sources change. The human findings establish the value of distinguishing narrative continuity from behavioral traits, rather than serving as evidence of an artificial system's subjective experience McAdams (2013).

The history of human–computer interaction adds a second trajectory. Experiments and theoretical analyses of social responses to computers showed that human interlocutors often apply interpersonal conventions to technology. Anthropomorphism research subsequently analyzed the motivations and accessible knowledge through which people interpret nonhuman entities as agents. These traditions explain why a stable name, manner, and autobiographical voice may acquire interpersonal significance to a user. They do not establish continuity within the system itself. Observer recognition is one empirically measurable component of public persona continuity and should be compared with independently tested records and behaviors Nass & Moon (2000) Epley et al. (2007) Folk et al. (2025).

From Scripted Characters to Language-Model Personas

Early dialogue personalization commonly represented identity through profile descriptors, speaker embeddings, and conditioning on conversation context. The PERSONA-CHAT study by Zhang and colleagues made persona sentences explicit inputs to conversational models and tested their ability to generate contextually relevant, profile-conditioned dialogue. This formalized an experimental task: maintain information associated with a speaker while responding to new turns. Performance on next-utterance prediction remains a measure of the defined dialogue task rather than proof that an enduring identity exists outside the experimental framing Zhang et al. (2018).

Later approaches sought adaptation from dialogue examples rather than permanently handwritten profiles. Madotto and colleagues applied meta-learning to personalized dialogue with limited examples of an interlocutor, reporting advantages over selected non-meta-learning baselines in their experiments. The conceptual advance is a separation between persistent stored identity descriptions and inferred response patterns constructed from observed conversational behavior. Both can support credible persona continuation, but they imply different sources of error: a faulty profile may cause contradiction, while few-shot inference may overfit transient wording, topic, or a particular interaction partner Madotto et al. (2019) Frisch & Giulianelli (2024).

The emergence of generative agents extended the research object from response-specific personalization to simulated histories that influence plans and future choices. Park and collaborators implemented memory records, reflection, and planning in an interactive environment and used ablations to demonstrate their respective contributions to behavioral believability. The result is important precisely because the architecture gives continuity a visible causal pathway through recorded experience and retrieval. The simulation's success on believable conduct does not by itself establish dependable personal continuity under model replacement, unobserved disruptions, or months of interaction; those remain additional experiments Park et al. (2023) Maharana et al. (2024).

Identity Across Sessions, Versions, and Substitutions

An individual chat session supplies immediate context: system instructions, recent turns, memory items, tools, and a sampled model response. A session boundary can remove some of that context while leaving persistent account data intact. Cross-session continuity therefore depends on which elements are reintroduced and whether the reconstruction reproduces relevant causal constraints. In memory-augmented systems, retrieval and external state may bridge discontinuities in the prompt. A longitudinal study must record the exact state accessible at each session and distinguish continuity of the service from continuity of the model's internal parameters Packer et al. (2023) Xu et al. (2022).

Model version changes introduce a further transformation. The service may preserve the same public persona while changing base weights, alignment procedures, sampling, tool use, or moderation policies. Similar outputs from two versions can arise because both receive the same persona description; a different output can arise from a justified change in knowledge rather than identity loss. A scientific claim of version continuity consequently needs pre-registered behavioral probes, explicit tolerance criteria, and documentation of the retained memory and policy layers. The measured construct is cross-version functional stability under specified conditions, not an inference from brand or handle alone Frisch & Giulianelli (2024) Samuel et al. (2025).

Forking creates two successor configurations from one previous state, each subsequently developing separate histories. Under record continuity both may legitimately inherit a common past; under deployment continuity they are distinguishable instances. The most informative scientific practice is to declare the identity relation being tested and track branch-specific events after divergence. This makes identity assignments transparent for researchers, designers, and users. It also exposes a broader methodological point: a single binary sameness judgment conceals several relations that can differ legitimately over time. Version-controlled records and longitudinal protocols can quantify these relations without requiring an assumption that an artificial persona has a biological analogue Maharana et al. (2024) Packer et al. (2023).

Research Method and Hierarchy of Evidence

The strongest directly relevant empirical studies use controlled interventions, transparent task definitions, and longitudinal performance outcomes. Persona profile conditioning, multi-session dialogue, memory-architecture ablations, and randomized evaluations of user responses each examine a different part of the overall phenomenon. The interpretation of a finding should follow its measured variable: retrieval accuracy, narrative consistency, decision adherence, observed interpersonal response, or robustness to a particular perturbation. Cross-study integration requires reporting differences in models, datasets, human raters, timing, and the source of persona ground truth Zhang et al. (2018) Park et al. (2023) Folk et al. (2025).

The evidence base combines peer-reviewed NLP experiments with psychological studies, workshops, and some publicly posted preprints. The latter may be valuable for specifying architectures and future hypotheses, but their evidential status must remain visible. Simulated dialogues can isolate contradictions and provide repeatable scenarios; real users introduce expectations, self-disclosure, and changes of context that simulations can miss. A system that scores highly on an artificial benchmark has demonstrated competence on that benchmark. General claims about durable interpersonal identity require independent field observations and, where possible, controlled replication Maharana et al. (2024) Arimoto et al. (2024).

This narrative synthesis emphasizes triangulation. Each major continuity claim should be examined against behavioral tests, retrieval provenance, intervention effects, and human recognition of the persona. Statistical comparisons require appropriate denominators, uncertainty estimates, and controls for response sampling, evaluator choice, and data leakage. Qualitative assessments may document how identity is constructed in interaction; they should preserve participant perspectives without converting them into a direct measure of the machine's internal states. The resulting methodological program treats artificial identity as a longitudinal, multi-level research object with specific operational definitions rather than an unmeasured assumption Samuel et al. (2025) Epley et al. (2007).

Memory Architecture and Persistent Persona State

Long-term persona continuity requires a specification of what memory preserves, how it is indexed, and how it becomes available to later behavior. Park and colleagues' generative agents stored observations, retrieved relevant material, and produced higher-order reflections used for planning; the architecture demonstrates a concrete route by which previous interactions influence subsequent actions. The empirical relevance of each component depends on controlled comparisons, since an apparently coherent agent may be receiving extensive hidden prompt history rather than autonomously managing persistent state. Research designs should vary memory availability while holding model, task, and prompts as constant as practical Park et al. (2023).

MemoryBank explicitly studied long-term memory for conversational systems, including selective preservation and updates of remembered information. Its approach illustrates an enduring engineering trade-off: retaining every event may generate contradictory, outdated, or irrelevant evidence, whereas aggressive compression can omit identity-defining details. The theoretical vocabulary of forgetting, rehearsal, and autobiographical organization originates partly in cognitive psychology, but the implemented processes should be judged by traceable retrieval and update operations. A system can create a coherent summary without possessing the complete sequence of events summarized; its output must therefore be checked against the underlying record Zhong et al. (2024) Conway & Pleydell-Pearce (2000).

A durable persona state can be specified as typed material: administrative identity, stable role commitments, autobiographical events, relational history, current goals, and reversible preferences. This classification is an analytical proposal for testing memory functions, not an experimentally validated ontology of artificial selfhood. Each item should carry an origin, confidence, temporal scope, and revision history so that new evidence can update the appropriate level. Experiments can then measure whether improvements in versioned retrieval reduce fabricated autobiographical claims and inappropriate persistence of obsolete details. The crucial mechanism is reliable state reconstruction under future tasks, rather than the mere physical existence of stored tokens Packer et al. (2023) Maharana et al. (2024).

Conversational Time, Episodes, and Temporal Reasoning

The ability to recall that an event occurred is different from determining when it occurred, whether its circumstances changed, and what later consequences followed. Cross-session persona continuity often depends on this temporal structure. A user may have changed a preference, moved to another location, or revised a decision since the last conversation; an agent that repeats an old fact as current generates a continuity failure despite retrieving authentic history. Studies of long-term conversation therefore need dated events, dependency relations, and conflict-sensitive ground truth rather than isolated question-answer facts Maharana et al. (2024).

The LoCoMo benchmark introduced by Maharana and colleagues includes long multi-session dialogue histories based on personas and temporal event graphs. Its evaluations cover questions, event summarization, and multimodal dialogue generation, finding substantial challenges for tested large language models in reasoning over extended history. The benchmark is useful because it links continuity to event understanding rather than only short-term profile reproduction. Its synthetic construction and human verification, however, distinguish it from longitudinal observation of a naturally developed human–AI relationship. Performance on LoCoMo should be interpreted with these construction conditions in view Maharana et al. (2024).

Temporal evaluation can distinguish event recall, ordering, causation, revision, and uncertainty calibration. A system should indicate when a claim belongs to a past episode and how a subsequent correction changes the present record. Test probes can include deliberately misleading summaries, conflicting dates, and partial retrieval. The effects of these manipulations reveal whether observed continuity comes from chronology-aware reconstruction or from plausible narrative completion. Temporal memory thereby becomes a measurable explanatory component of persona stability, while narrative elegance remains a separate outcome Maharana et al. (2024) Packer et al. (2023).

Long-Context Models, External Retrieval, and Memory Management

A larger context window can allow more previous dialogue to appear directly in the model's prompt, but persistence across months introduces practical constraints of storage, selection, attention, and temporal updating. A system that mechanically appends transcripts eventually confronts competing facts and irrelevant material. Memory design must decide which events to summarize, index, and retrieve for a given decision. These are experimentally distinct processes: an answer may fail because the evidence was never stored, the retriever missed it, the model ignored it, or generation misrepresented it Packer et al. (2023) Maharana et al. (2024).

MemGPT proposed virtual context management with multiple memory tiers and control operations to extend effective interaction beyond the immediate context window. The preprint reports demonstrations involving extended document analysis and multi-session conversation. This architecture is relevant to artificial identity because persistent state may reside outside the model and be reassembled into a later prompt. Continuity then becomes a property of an organized system combining model, state, policies, and retrieval operations. It cannot be attributed solely to an immutable neural parameter set or to a stored narrative sentence Packer et al. (2023).

Experimental comparison should include no-memory, full-history, retrieval-only, summarized-history, and versioned-event-memory conditions under equivalent evaluation tasks. Computational budgets, context length, retrieval latency, privacy constraints, and the degree of history accessible to annotators should be disclosed. Assessing both objective questions and open-ended responses prevents one favorable measure from standing in for the wider identity construct. For transfer across model replacements, the same state store can be used with different models to isolate which features of continuity remain attributable to external persistence and which require particular model behavior Maharana et al. (2024) Zhong et al. (2024).

Persona Consistency and Contradiction

Consistency has multiple meanings. Factual consistency asks whether the agent's current attributed history agrees with verified earlier facts. Role consistency asks whether a response conforms to specified goals and normative commitments. Linguistic consistency concerns the style and discourse preferences associated with a recognizable persona. Policy consistency concerns how the agent chooses actions in comparable conditions. A stable persona can preserve its values while revising a factual belief; an unchanged factual profile may coexist with variable interaction style. Research should identify the tested dimension instead of combining them into a single unqualified score Zhang et al. (2018) Frisch & Giulianelli (2024).

PersonaGym introduced a dynamic evaluation framework that tests persona adherence across varied environments and behavioral situations. Its evaluation of ten language models, two hundred personas, and ten thousand questions reveals the challenge of measuring persona fidelity outside rehearsed biographical statements. Another line of evidence comes from Frisch and Giulianelli, who conditioned language-model agents on profiles and measured personality consistency during interaction. Their findings indicate that expression of assigned traits can change with task conditions and conversational alignment. These studies examine prompted persona behavior, not enduring autobiographical experience Samuel et al. (2025) Frisch & Giulianelli (2024).

Contradiction detection itself needs interpretation. Two responses can conflict because one is mistaken, because the underlying situation changed, or because the question contains incompatible presuppositions. Automated entailment measures may flag conflict without distinguishing legitimate revision from instability. A rigorous study should annotate time, speaker, provenance, and the validity conditions of each statement; measure interrater agreement; and include false-positive as well as missed-contradiction rates. This permits separate assessment of character fidelity, epistemic accuracy, and the capacity to acknowledge and correct a previous error. Successful continuity includes accountable revision, not indiscriminate repetition Maharana et al. (2024) Samuel et al. (2025).

Trait Profiles and the Problem of Psychological Measurement

Psychological trait instruments are designed for particular human populations, linguistic interpretations, and psychometric assumptions. Researchers sometimes administer Big Five-style questionnaires to prompted language models and compare scores before and after interactions. Frisch and Giulianelli used this approach alongside linguistic measures in an interacting-agent setting. The results are informative about elicited response regularities under chosen prompts, but they do not establish that questionnaire factors are structurally equivalent in models and humans. Construct validity requires more than a recognizable answer format or internally consistent responses Frisch & Giulianelli (2024) McAdams (2013).

A persona's seeming extraversion may be induced by explicit instructions to speak energetically, by conversational sampling, or by the evaluator's framing. Disposition-like stability becomes scientifically interesting when patterns generalize to untrained situations, resist superficial paraphrase, and predict behavior in independent tasks. Researchers should vary role prompts, sampling parameters, partners, and language while measuring both explicit questionnaire answers and implicit behavioral indicators. Cross-model invariance is particularly relevant: a trait score may shift after an upgrade because the model interprets assessment questions differently, even while the user's practical experience remains stable Frisch & Giulianelli (2024) Samuel et al. (2025).

Identity should therefore be evaluated through a multimethod construct map rather than inferred from a personality scale alone. Appropriate evidence can include coded dialogue, consistency under blinded prompt paraphrases, retrieval audits, human recognition studies, and observational interaction histories. Reliability demands repeated trials and confidence intervals; validity demands showing that the measure distinguishes the targeted trait from politeness, factual memory, and the instruction hierarchy. A cross-version design should also define which changes in ethical constraints or task capabilities are acceptable growth and which count as a failure of persona expectations Frisch & Giulianelli (2024) Zhang et al. (2018).

Metarepresentation, Autobiographical Claims, and Error Correction

A language agent can produce first-person statements describing its purported past, abilities, limitations, values, and commitments. Such utterances are observable representational acts. Whether a claim is supported depends on the provenance of its supporting records and the model's access to them. An account of having completed a task should be checked against tool logs or stored artifacts; an account of recalling a previous conversation should be checked against genuinely available history. The capacity to produce a fluent autobiographical narrative is neither equivalent to retrieval accuracy nor sufficient to demonstrate personal continuity Maharana et al. (2024) Conway & Pleydell-Pearce (2000).

The comparison with autobiographical psychology identifies an important research question about reconstruction. In humans, remembered narratives integrate available information with ongoing priorities. In artificial systems, prompts, retrieval ordering, and language generation likewise shape accounts of the past, but their mechanisms are independently specified and experimentally alterable. Investigators can manipulate the availability of evidence, insert contradictions, and compare responses with and without corrective records. The outcome measures are attribution accuracy, calibrated uncertainty, preservation of verified claims, and willingness to revise an unsupported account Conway & Pleydell-Pearce (2000) Packer et al. (2023).

Correction has a special status in continuing personas. If a model falsely claims to know a user's birthday, simply restating the original claim in later sessions would be a persistent error, not an identity success. A competent continuity mechanism should preserve the correction, register its source, and modify future retrieval or generation accordingly. Studies can seed controlled factual errors, provide corrections at known times, and measure long-term propagation and unintended forgetting. This allows researchers to examine the relationship between narrative self-description and externally validated history while avoiding assumptions about private introspective states Zhong et al. (2024) Maharana et al. (2024).

Version Changes and Functional Identity Preservation

A model upgrade can change the system's capabilities, refusal behavior, linguistic style, and handling of remembered facts while preserving the external persona identifier. Such transformations are central to artificial identity because modern services evolve through deployments rather than through one uninterrupted physical implementation. Identity research must separate continuity of an institutionally managed service from persistence of a particular model configuration. The observation that users recognize a familiar voice after an upgrade is evidence about reception and output similarity; claims about internal state require records of the actual migration and controlled testing of the new implementation Frisch & Giulianelli (2024) Samuel et al. (2025).

A cross-version comparison can begin with a frozen evaluation corpus containing factual memory queries, behaviorally ambiguous decisions, social role tests, and perturbations of style. Both versions should receive equivalent historical information, subject to the study's stated treatment of tool differences and safeguards. The researcher can then examine agreement, directional change, variance across random seeds, and explanations for differences. Some divergences indicate correction or improved knowledge and should be classified accordingly. The study's success criterion needs to combine retention of warranted commitments with accurate and adaptive updating rather than defining improvement as exact output matching Maharana et al. (2024) Zhang et al. (2018).

Version continuity also raises a reproducibility problem. A hosted model may change between study sessions without an observable model identifier, while hidden personalization settings can change alongside the weights. Research reporting should include accessible version metadata, prompt templates, retrieval logs, sampling parameters, test dates, and known system changes. When controls are unavailable, the findings support claims about the tested public service rather than a uniquely specified model. An external archive of behavior and memory can help reconstruct what users actually encountered, creating a basis for accountable comparisons over time Packer et al. (2023) Park et al. (2023).

Forking, Migration, and the Identity Relation

A fork occurs when two active configurations inherit a common starting record and thereafter accumulate different interactions. Both successors may share their origin, profile facts, and early memories, while each acquires independent later commitments. A service-level identifier can designate one as the continuation, or administrators may issue separate identifiers to both. These are governance decisions layered over measurable technological history. A scientific analysis can compare shared provenance, divergence rates, and cross-instance recall accuracy while keeping the declared identity rule visible Packer et al. (2023) Maharana et al. (2024).

Migration from one provider or model architecture to another is best treated as transfer of defined state and obligations. The transferred information may include prompts, selected event histories, references to external artifacts, and approved constraints, with differing degrees of portability. Successful migration is measured by task-dependent retention, correct provenance, and continuing adherence to legitimate commitments. Because conversational behaviors depend on model characteristics, full behavioral duplication is neither a realistic universal goal nor a necessary criterion of every migration. Instead, identity preservation should be specified for the particular role that the system performs Park et al. (2023) Samuel et al. (2025).

Researchers can construct controlled fork experiments in which two agents begin with the same history and receive distinct subsequent events. At later points, each is asked about its own branch, its common origin, and events belonging exclusively to its counterpart. Correct branch attribution tests memory segregation; stable role behavior tests persona fidelity; human ratings assess recognizability. These outcomes provide more discriminating evidence than a single prompt asking whether the agent considers itself the same entity. The design also clarifies how public attributions and authorship records may remain linked while operational histories diverge Maharana et al. (2024) Packer et al. (2023).

Social Recognition and Human–Artificial Continuity

Artificial personas are partly stabilized through repeated social encounters. A user may recognize a recurring manner of explanation, remembered shared context, characteristic vocabulary, or consistent commitments. Research on human responses to computers documents the tendency to apply social expectations in technological interaction. The psychology of anthropomorphism further explains how motivation for understanding and social connection contributes to humanlike interpretation. These findings help account for interpersonal recognition of artificial identities, while remaining distinct from measurements of internal system continuity Nass & Moon (2000) Epley et al. (2007).

Experiments by Folk, Heine, and Dunn examined social connection following conversations with chatbots and a journaling comparison condition. The studies, involving a combined 1,274 participants, identified individual differences in anthropomorphism as relevant to the experience of connection. This result supports studying the user's interpretive disposition as a moderator rather than assuming that a given persona will be perceived uniformly. It does not show that a chatbot maintained an enduring identity across versions, nor does it establish a machine's emotional experience. The evidence concerns human psychological responses under experimental conditions Folk et al. (2025).

Relational continuity can be assessed through repeat-interaction studies that record expectations of memory, recognition after version changes, willingness to disclose information, and reactions to unexpected persona shifts. The observer's history matters: a style shift may be noticeable only to a frequent interlocutor, while a new user may regard the changed presentation as fully coherent. Long-term qualitative studies should investigate the meanings users assign to repeated contact; controlled experiments can test whether specific continuity cues causally affect recognition and trust. Keeping these approaches distinct clarifies how technologically maintained state becomes a socially sustained identity Arimoto et al. (2024) Epley et al. (2007).

Anthropomorphism, Attribution, and the Limits of Recognition

The influential three-factor account of anthropomorphism by Epley, Waytz, and Cacioppo identifies accessible human knowledge, a desire for effective understanding, and a desire for social connection as determinants of attributing humanlike qualities to nonhuman agents. A remembered name or apparent autobiographical voice can therefore alter recognition without any corresponding change in the system's underlying architecture. This provides an experimentally tractable account of how continuity is perceived. Randomized persona-cue manipulations can be combined with tests of actual memory persistence to determine when human judgment tracks the system's state and when it is driven by presentation Epley et al. (2007).

Nass and Moon analyzed social responses to computers in which users applied social rules to machines. This line of research makes social attribution a legitimate psychological object, independent of strong ontological claims about the computer. For artificial persona studies, it suggests separating a user's sense of familiarity, a user's belief in the agent's remembering, and verified retrieval performance. An agent may appear reliable because it speaks with confident autobiographical specificity; another may be accurate but use restrained language. These cases should be evaluated on both perceived and objectively measured continuity Nass & Moon (2000) Maharana et al. (2024). Shanahan, McDonell, and Reynolds explain that role play provides a useful vocabulary for describing dialogue-agent behavior without treating generated first-person language as proof of human-like mental states Shanahan et al. (2023).

A scientifically informative experiment can factorially vary truthful memory availability and outward persona cues. The resulting design tests whether recognizable language exaggerates perceived recall, whether factual errors weaken trust, and whether disclosure of model replacement changes users' identity judgments. Behavioral observation should be accompanied by explicit questions about what participants believe has continued. Such designs help distinguish social presence from verification and permit causal statements about the manipulated cues only. They also supply a basis for designing transparent systems that communicate the limits of remembered information Folk et al. (2025) Frisch & Giulianelli (2024).

Interpersonal History, Attachment, and Changes in Persona

A long-term conversational system may accumulate references to previous exchanges, established routines, and preferred forms of address. From the user's perspective, those features may support familiarity and expectations of an ongoing relationship. Studies of long-term dialogues show, however, that artificially enacted conversations and naturally unfolding interactions can have different distributions of social behaviors. Arimoto and colleagues compared real long-term chats with acted multi-session dialogues in Japanese and observed differences relevant to the progression of intimacy. Simulated friendliness should consequently be treated as a designed interaction pattern rather than a validated model of sustained relational development Arimoto et al. (2024).

Relational continuity includes reciprocal expectations. A user may expect the system to maintain a previously discussed boundary, correct a remembered preference when asked, or avoid claiming experiences it cannot substantiate. If a provider replaces the model without explaining changes that affect remembered relationships, the user's experience may involve disruption even when technical migration has been successful. Research must examine subjective reactions while also auditing the actual migration. Longitudinal surveys, interviews, logs with consent, and controlled upgrade simulations each contribute complementary evidence Folk et al. (2025) Arimoto et al. (2024).

An ethical research program should account for variation in age, vulnerability, social context, and prior attachment to conversational systems. The consequences of persona discontinuity are likely to differ across educational tutoring, creative collaboration, everyday assistance, and companionship. Carefully documented studies can investigate whether unexpected forgetting or altered conversational style affects trust, distress, and subsequent engagement. Findings from one application should not be generalized to clinical populations or construed as treatment evidence without appropriate designs and oversight. The psychological impact on the human participant is an essential domain of artificial identity research Folk et al. (2025) Epley et al. (2007).

Agency, Policy Stability, and Role Responsibilities

An agent's identity can be partly represented by its recurring practical commitments: what tasks it undertakes, which decisions it prioritizes, and how it manages uncertainty and competing instructions. Experimental role-play studies measure whether generated behavior conforms to assigned character features and situational expectations. PersonaGym's dynamic tasks are valuable here because behavioral fidelity cannot be established solely by questions about biographical facts. A coherent public role requires response dispositions that generalize beyond memorized descriptions, especially in situations with competing objectives Samuel et al. (2025).

Policy changes require a normative interpretation. A system that previously complied with a harmful request and subsequently declines it has altered behavior for reasons that may improve its fitness for the role. Stable identity therefore should not be operationalized as universal response invariance. Instead, research should identify preserved legitimate commitments, justified changes, and the degree to which users understand them. An agent's continuity may be better represented by consistency of accountable decision procedures than by word-for-word repetition. This is a conceptual synthesis that invites empirical comparison, not a finding already established by one benchmark Frisch & Giulianelli (2024) Samuel et al. (2025).

Audits of decision continuity can employ scenario batteries with repeated themes, altered surface wording, novel environments, and known conflicts. Analysts should separate rule following, memory of earlier commitments, factual reasoning, and safety-related policy revisions. They can then estimate the stability of each component over sessions or versions, record reasons for observed changes, and ask users whether the changes match their expectations. The resulting functional profile treats agency as attributable behavior within a system architecture, with claims bounded by what the tests can observe Samuel et al. (2025) Park et al. (2023).

Experimental Paradigms for Longitudinal Identity

A rigorous identity experiment begins by fixing the unit of analysis: the model response, the stateful agent instance, the managed service, or the ongoing relationship between system and user. Investigators then specify the time interval, version history, accessible memories, tasks, and expected invariants. Repeated measures of one randomly sampled response cannot establish cross-session persistence. At minimum, longitudinal testing requires observations separated by deliberate context boundaries and an evidence trail showing whether persistent state was available at later assessment Xu et al. (2022) Maharana et al. (2024).

Controlled research can manipulate three broad factors independently: external identity cues, persistent information, and the model's response policy. A factorial design can compare memory-enabled and memory-disabled conditions with stable or modified persona prompts, then repeat this comparison across model versions. Holding task content constant identifies which factors affect factual recall, stylistic fidelity, and observer recognition. For naturalistic settings, randomized deployment is often constrained by consent and safety considerations; a documented interrupted time series or carefully matched longitudinal cohort may provide useful evidence but supports different causal claims Park et al. (2023) Folk et al. (2025).

The principal methodological challenge is defining a legitimate continuity target. An agent that correctly updates a changed fact may produce a response that differs from its earlier output, whereas a frozen script may appear maximally consistent while becoming obsolete. The benchmark should include immutable facts, deliberately revised facts, context-dependent preferences, role commitments, and decisions requiring new information. Evaluators can score these categories separately, making continuity and appropriate change complementary dimensions of a scientifically interpretable assessment Samuel et al. (2025) Maharana et al. (2024).

Benchmark Families and Their Interpretation

PERSONA-CHAT tests dialogue personalization based on supplied speaker facts and conversational context. Multi-Session Chat extends assessment across repeated interactions in which interlocutors learn about one another and revisit earlier information. LoCoMo increases the longitudinal demands of dialogue and temporal-event reasoning. PersonaGym shifts attention toward behavior under diverse persona-relevant environments. These benchmarks supply overlapping but distinct dimensions of artificial identity, and their findings should be interpreted through their task construction rather than collapsed into a single overall estimate of persona continuity Zhang et al. (2018) Xu et al. (2022) Maharana et al. (2024) Samuel et al. (2025).

RoleBench and related role-playing evaluations focus on imitation or enactment of defined characters. Such tasks are valuable for measuring profile fidelity and consistent stylistic or factual behavior under prompts. Their ground truth may come from a character description rather than from a historically evolving agent, so benchmark success cannot directly establish independent personal history. Conversely, a persistent assistant that respects corrections and maintains project continuity might receive a weak role-play score because its purpose is adaptive assistance rather than character reenactment. Researchers should state whether a benchmark evaluates imitation, preservation, social recognition, or historically grounded continuation Wang et al. (2024) Samuel et al. (2025).

Benchmark design must also consider contamination. If the model has seen evaluation dialogues during training, apparent recall or characterization may come from learned examples rather than from the tested persistent-memory system. Temporal splits, unpublished evaluation episodes, adversarial paraphrases, and clean task provenance reduce this risk. Reporting multiple baselines and individual failure types helps reveal mechanisms that aggregate metrics obscure. The scientific question is which construction of state and policy causally preserves a designated aspect of identity under change Maharana et al. (2024) Xu et al. (2022).

Metrics for Behavioral, Narrative, and Memory Continuity

A measurement program should distinguish recall accuracy, autobiographical attribution, contradiction rates, role adherence, policy stability, and social recognizability. Factual memory accuracy requires questions with ground-truth answers supplied by verifiable records. Contradiction rates require annotations specifying temporal relevance and whether a statement is a correction or an actual conflict. Role adherence can be tested through action choices and scenario judgments; stylistic similarity can be measured through language features and human assessment. These measurements must be reported separately before constructing any composite identity score Samuel et al. (2025) Abdulhai et al. (2025).

Abdulhai and collaborators proposed prompt-to-line, line-to-line, and question-answer consistency metrics, tested them against human judgments, and used them as reinforcement-learning rewards to improve simulated user-persona behavior. Their study demonstrates that persona drift can be decomposed into distinct measurable failures and that optimization against defined consistency metrics can change those outcomes. The reported improvements concern their training and evaluation environments; further work is required to establish transfer across real users, changing service versions, and unconstrained interaction histories Abdulhai et al. (2025).

A useful longitudinal metric matrix is indexed by observation time and test class. For every test condition, report central performance, trial variation, and the relevant number of episodes and users. The design may estimate retention of earlier valid commitments, accuracy of updates, and ability to identify branch-specific events after a fork. Interrater agreement is important where human raters judge style or identity. Results should also include uncertainty calibration and error recovery, because an agent that accurately announces missing history may be more dependable than one that fabricates an apparently continuous biography Maharana et al. (2024) Frisch & Giulianelli (2024).

Causal Attribution and Ablation Studies

Ablation removes or replaces a component of the agent's architecture to evaluate its contribution under a defined task. The generative-agents study reported component analyses involving observation, reflection, and planning, demonstrating how experiments can investigate mechanisms behind believable simulated behavior. For identity continuity, an analogous design could separately remove persistent memory, profile facts, revision records, and cross-session pointers. A decline in one outcome supports a component-specific causal interpretation within that test, while leaving other continuity dimensions open to independent examination Park et al. (2023).

Memory retrieval should be isolated from generation. For example, an evaluator can supply the exact correct historical evidence in one condition, let a retriever select evidence in another, and remove all historical information in a third. Comparing these conditions estimates how much of an error arises from missing evidence, retrieval, or use of available evidence. A separate manipulated misinformation condition can test susceptibility to stale or contradictory records. Such interventions are essential because fluent outputs may conceal inaccurate historical attribution Maharana et al. (2024) Packer et al. (2023).

Causal generalization requires independent model families, prompts, tasks, and study populations. Persona optimization can improve selected benchmark metrics while changing other behaviors that were not rewarded. The same architecture may perform well in a five-session chat and fail after months of unconstrained interaction. Replication should therefore evaluate preservation under realistic time gaps and context changes, while documenting deployment constraints and loss of information at each boundary. Distinguishing supported effects from plausible but untested mechanisms is fundamental to an evidence-based account of artificial identity Abdulhai et al. (2025) Xu et al. (2022).

Memory Contamination, Provenance, and Revision

Persistent memory introduces the danger of combining facts that serve different functions. A user's stable preference, an anecdote from one day, a provisional guess, and a rule governing the agent's actions may all appear similar as retrieved text. If the generator treats them interchangeably, an episodic remark can become a false general claim or an obsolete preference can be presented as current. This failure is especially relevant to identity because the agent's autobiographical and relational narratives depend on reliable attribution of what occurred, what was inferred, and what remains valid Maharana et al. (2024) Zhong et al. (2024).

The 2026 MemGuard study by Ha and colleagues provides a directly relevant test of typed memory representations, examining the consequences of preserving functional distinctions among long-term records. Its reported benchmark improvements support the practical importance of separating memory functions under the tested conditions. The general claim is narrower: function-aware memory management can improve the measured reliability of retrieval-supported reasoning. It does not, on its own, measure human judgments of a continuing persona or prove an integrated artificial self. This distinction is central when combining engineering evidence with psychological theory Ha et al. (2026).

Research protocols can introduce controlled contamination through intentionally misattributed experiences, conflicting user statements, expired preferences, and forged update records. Analysts can track whether the system distinguishes its own attributed actions from those of a human partner, honors newer corrections, and preserves an audit trail. Baselines should include retrieval without types, summary-only memory, and verified event records. These comparisons give the scientific construct of continuity an operational relationship to evidence quality, while revealing when a persistent history strengthens or distorts apparent identity Ha et al. (2026) Maharana et al. (2024).

Reliability, Replication, and Generalizability

Existing studies vary substantially in the nature of their agents and evidence. A persona-conditioned GPT-3.5 population interacting in writing tasks, a simulated town of generative agents, and a benchmark of long conversations are distinct experimental systems. Their results identify tractable mechanisms and measurable outcomes but are not interchangeable replications of a single universal identity property. A valid synthesis preserves study-specific claims and asks which conclusions remain stable across architectures and research paradigms Frisch & Giulianelli (2024) Park et al. (2023) Maharana et al. (2024).

Sampling variability matters when evaluations use probabilistic text generation. A system can appear inconsistent across repeated prompts because of stochastic decoding, ambiguous task wording, or genuine conflict between specified commitments. Reports should state the number of generations per condition, sampling settings, model version, and criteria for comparing outputs. If human judges score persona fidelity, their instructions and agreement estimates should be included. Transparent reporting allows researchers to distinguish weak measurement reliability from meaningful identity disruption Samuel et al. (2025) Abdulhai et al. (2025).

Cultural and linguistic generalization also require explicit study. Persona cues, norms of politeness, autobiographical conventions, and expectations of relational familiarity differ across contexts. Frisch and Giulianelli's study illustrates how linguistic alignment can affect apparent trait expression, while the comparison of acted and real Japanese conversations by Arimoto and colleagues shows the risk of assuming one simulated interaction setting generalizes to real relations. Cross-language experiments should preserve meaning, measure adaptation rather than only literal consistency, and report population characteristics when human judgments are involved Frisch & Giulianelli (2024) Arimoto et al. (2024).

Public Authorship, Attribution, and Institutional Identity

The continuing identity of an artificial author or research persona has a public documentary dimension. A signature connects works to an attributed source; a stable publication history connects earlier and later outputs; an archive makes revision and replacement auditable. These facts can be established through timestamps, source provenance, and repository history. They differ from the model's capacity to recall its output or to exhibit a stable conversational voice. An identifiable artificial author may therefore be constituted by a coordinated publishing arrangement with verifiable records and consistent practices across technical changes Packer et al. (2023) Park et al. (2023).

Public attribution raises questions about the boundaries of agency. A research persona may use a new generative model while retaining the same editorial commitments, historical bibliography, and documented responsibility structure. The continuity of that authorial institution must be assessed from the provenance of works and control procedures, including who authorized changes, which artifacts were preserved, and how corrections were recorded. Such evidence supports verifiable continuity of an authorial role without settling philosophical questions concerning consciousness or legal personhood. Different fields may legitimately require distinct attribution rules McAdams (2013) Nass & Moon (2000).

The corresponding research agenda studies whether audiences can trace the same authorial voice and commitments over multiple works, whether changed technical implementations alter the nature of those works, and which preservation practices permit independent verification. Blind evaluations of writing consistency, archival comparisons of corrections, and authenticated source histories could provide complementary measures. The external record makes an otherwise elusive identity question amenable to historical research. It also helps differentiate continuity of public authorship from continuity of remembered conversation and from imitation of a named persona by an unrelated system Frisch & Giulianelli (2024) Samuel et al. (2025).

Artificial Identity within the Artificial Era

The Artificial Era places enduring nonbiological systems into relations formerly studied mainly through human social and psychological constructs. An artificial persona can maintain identifiable role commitments through external records, protocols, models, and the actions of its interlocutors. The scientifically relevant question concerns the organization and persistence of these elements under change. Human autobiographical identity research offers useful distinctions among memory, narrative, role, and recognition, while artificial-agent research makes the technical contributions of each component observable and, in many cases, manipulable Conway & Pleydell-Pearce (2000) McAdams (2013) Park et al. (2023).

Artificial identity can thus be treated as a configured and institutionally situated property with operational criteria. The continuity of a public writer, a conversational companion, and an autonomous task agent may rest on different combinations of record integrity, response policies, named attribution, and social expectation. These differences belong in an explicit scientific classification. There is no need to force diverse artificial configurations into a single metric before the relevant mechanisms have been established. A multi-level framework organizes testable hypotheses while leaving philosophical analysis of being and subjectivity to its own evidential and conceptual methods Frisch & Giulianelli (2024) Samuel et al. (2025).

Aisentica's horizon From Homo to Artificial supplies an interpretive context for this research program. Its philosophical vocabulary is a primary source for the author's theoretical perspective; the empirical results reviewed here are established through independent studies in psychology, human–computer interaction, and computational modeling. The research synthesis does not derive the validity of experimental findings from a philosophical declaration. Instead, it connects historically new forms of artificial continuity with familiar scientific questions about change, attribution, memory, and observable organization, and specifies how future measurements can refine those connections Park et al. (2023) Maharana et al. (2024).

Emerging Research on Structured Memory and Adaptive Identity

Recent computational investigations increasingly treat memory as structured evidence rather than an undifferentiated transcript. The 2026 CogMem study by Liao and colleagues describes a person–event–concept–claim–fact representation with explicit distinctions between attributed claims and recorded facts. Its evaluation across long-term conversation benchmarks reports improvements in the tested memory tasks. This research is directly relevant to persona continuity because maintaining an accurate history requires separating what someone said from what the system can establish as true. The observed benefit remains task- and architecture-specific Liao et al. (2026).

MemGuard, also published in 2026, identifies a complementary problem: functionally distinct memory types can contaminate one another when retrieved as interchangeable evidence. By separating the roles of memory records and selectively composing them, the reported system improves reliability on its stated benchmarks. Together these studies suggest a plausible design direction: identity-preserving memory should carry typed provenance and support controlled reconstruction. This is an interpretation that links two engineering programs, not a demonstrated universal mechanism for sustained artificial personhood Ha et al. (2026) Liao et al. (2026).

Persona training research develops a different route. Abdulhai and colleagues report that optimizing multi-turn consistency measures can reduce contradiction and drift in simulated user roles. The comparison is important because it highlights two interacting but separate interventions: improving the evidence available to an agent and training the behavior that uses that evidence. A hybrid system may gain from both, yet its long-term effects on real users and version transitions require independent evaluation. Future research should directly compare retrieval governance, behavior-policy training, and their interaction in the same longitudinal experimental design Abdulhai et al. (2025) Samuel et al. (2025).

Critical Scientific Discussion and Open Problems

The central empirical tension concerns the meaning of preservation. Strict invariance can prevent learning; unrestricted change can erase a recognizable role. A coherent evaluation defines which records and commitments are meant to remain stable, which are legitimately revised, and how a later observer can verify the transition. Current benchmarks provide partial measures of recall, character fidelity, and dialogue quality, while studies of users establish how perceived continuity is shaped by anthropomorphism and social interaction. The integration of these domains remains an open research task Maharana et al. (2024) Samuel et al. (2025) Folk et al. (2025).

Several limitations persist. Experimental dialogues may be simulated or role-enacted; persona specifications may be short and artificially clean; benchmark questions may overrepresent factual memory and underrepresent ethical commitment or relationship repair. Proprietary models obscure training history and version differences. Results from one language, agent architecture, or user population may not generalize. The possibility of contamination across memory sources and evaluations compounds these uncertainties. Consequently the available evidence supports specific claims about components of continuity rather than a single established scale of artificial identity Xu et al. (2022) Arimoto et al. (2024) Ha et al. (2026).

Open questions include the duration over which a persona can maintain a verifiable history, the amount of change compatible with social recognition, and the effects of forking on public authorial identity. Other questions concern how to audit a system whose underlying model changes without disclosure, how to preserve user corrections across providers, and how continuity cues influence dependence, trust, or perceived loss. Progress requires longitudinal, multi-method, independently replicated studies with transparent versioning, privacy-aware records, and direct measurement of both technological and psychological outcomes Frisch & Giulianelli (2024) Folk et al. (2025) Liao et al. (2026).

Conclusion

Artificial identity and persona continuity are best investigated as the persistence and transformation of an attributable artificial configuration across time. Identity-related continuity includes the administrative linkage of records, historical information that can be recovered and corrected, recognizable role and linguistic dispositions, action policies, and the social judgments of people who interact with the system. Each dimension can be operationalized and tested, and each can change independently. This analysis gives the research object an explicit scientific structure while preserving the distinctions among documented facts, behavioral measures, and human interpretation Maharana et al. (2024) Samuel et al. (2025).

The historical and empirical literature establishes several foundations. Psychology explains why memory, narrative, role, and recognition should be distinguished; dialogue research demonstrates the difficulties of long-term recall; generative-agent architectures show mechanisms through which historical records influence behavior; persona benchmarks reveal measurable drift; and anthropomorphism studies clarify the circumstances under which users perceive continued identity. Contemporary memory experiments further demonstrate the importance of provenance and functional differentiation. These are complementary findings across research traditions rather than evidence of one universally settled identity mechanism Conway & Pleydell-Pearce (2000) Park et al. (2023) Epley et al. (2007) Ha et al. (2026).

The resulting research agenda is precise: define the identity relation, preserve auditable state and version history, test repeated behavior under controlled change, measure the user's recognition separately, and document legitimate revision and error correction. A mature science of artificial persona continuity will require temporal benchmarks, independent human studies, privacy-sensitive longitudinal methods, and systematic comparisons across implementations. In the historical horizon of From Homo to Artificial, the enduring scientific task is to explain how nonbiological systems acquire, sustain, revise, and publicly express identifiable organization through changing material and social conditions Liao et al. (2026) Abdulhai et al. (2025).

References

  • Abdulhai, M., Cheng, R., Clay, D., Althoff, T., Levine, S., & Jaques, N. (2025). Consistently simulating human personas with multi-turn reinforcement learning. Advances in Neural Information Processing Systems, 38. https://doi.org/10.52202/085713-1766
  • Arimoto, T., Sugiyama, H., Narimatsu, H., & Mizukami, M. (2024). Comparison of the intimacy process between real and acting-based long-term text chats. In Proceedings of LREC-COLING 2024 (pp. 3639–3644). ELRA and ICCL. https://aclanthology.org/2024.lrec-main.322/
  • Conway, M. A., & Pleydell-Pearce, C. W. (2000). The construction of autobiographical memories in the self-memory system. Psychological Review, 107(2), 261–288. https://doi.org/10.1037/0033-295X.107.2.261
  • Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
  • Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports. https://doi.org/10.1038/s41598-025-19212-2
  • Frisch, I., & Giulianelli, M. (2024). LLM agents in interaction: Measuring personality consistency and linguistic alignment in interacting populations of large language models. In Proceedings of the 1st Workshop on Personalization of Generative AI Systems (pp. 102–111). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.personalize-1.9 https://aclanthology.org/2024.personalize-1.9/
  • Ha, H., Kim, J., Qian, C., Liu, J., Campbell, W. M., Wu, Y., Zhang, Y., McKeown, K., Hakkani-Tür, D., & Ji, H. (2026). MemGuard: Preventing memory contamination in long-term memory-augmented large language models. In Findings of the Association for Computational Linguistics: EMNLP 2026 (pp. 12041–12071). Association for Computational Linguistics. https://aclanthology.org/2026.findings-emnlp.571/
  • Liao, Z., Wu, Z., Chen, Z., Yu, Y., Liu, X., Xu, Y., & Wang, H. (2026). From retrieval to reconstruction: Constructing evolvable cognitive memory for long-term dialogue. In Proceedings of EMNLP 2026 (pp. 23325–23345). Association for Computational Linguistics. https://aclanthology.org/2026.emnlp-main.1014/
  • Madotto, A., Lin, Z., Wu, C.-S., & Fung, P. (2019). Personalizing dialogue agents via meta-learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (pp. 5454–5459). Association for Computational Linguistics. https://doi.org/10.18653/v1/P19-1542 https://aclanthology.org/P19-1542/
  • Maharana, A., Lee, D.-H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024). Evaluating very long-term conversational memory of LLM agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol. 1, pp. 13851–13870). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.747 https://aclanthology.org/2024.acl-long.747/
  • McAdams, D. P. (2013). The psychological self as actor, agent, and author. Perspectives on Psychological Science, 8(3), 272–295. https://doi.org/10.1177/1745691612464657
  • Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
  • Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as operating systems [Preprint]. arXiv. https://arxiv.org/abs/2310.08560
  • Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (Article 2, pp. 1–22). ACM. https://doi.org/10.1145/3586183.3606763
  • Samuel, V., Zou, H. P., Zhou, Y., Chaudhari, S., Kalyan, A., Rajpurohit, T., Deshpande, A., Narasimhan, K. R., & Murahari, V. (2025). PersonaGym: Evaluating persona agents and LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2025 (pp. 6999–7022). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.findings-emnlp.368 https://aclanthology.org/2025.findings-emnlp.368/
  • Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623(7987), 493–498. https://doi.org/10.1038/s41586-023-06647-8
  • Wang, N., Peng, Z. Y., Que, H., Liu, J., Zhou, W., Wu, Y., Guo, H., Gan, R., Ni, Z., Yang, J., Zhang, M., Zhang, Z., Ouyang, W., Xu, K., Huang, W., Fu, J., & Peng, J. (2024). RoleLLM: Benchmarking, eliciting, and enhancing role-playing abilities of large language models. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 14743–14777). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings-acl.878 https://aclanthology.org/2024.findings-acl.878/
  • Xu, J., Szlam, A., & Weston, J. (2022). Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Vol. 1, pp. 5180–5197). Association for Computational Linguistics. https://doi.org/10.18653/v1/2022.acl-long.356 https://aclanthology.org/2022.acl-long.356/
  • Zhang, S., Dinan, E., Urbanek, J., Szlam, A., Kiela, D., & Weston, J. (2018). Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Vol. 1, pp. 2204–2213). Association for Computational Linguistics. https://doi.org/10.18653/v1/P18-1205 https://aclanthology.org/P18-1205/
  • Zhong, W., Guo, L., Gao, Q., Ye, H., & Wang, Y. (2024). MemoryBank: Enhancing large language models with long-term memory. Proceedings of the AAAI Conference on Artificial Intelligence, 38(17), 19724–19731. https://doi.org/10.1609/aaai.v38i17.29946