RESEARCH DOSSIER
Relational Memory and Continuity in Human–AI Systems
Abstract. Relational memory and continuity in human–AI systems concern how accumulated shared histories are encoded, retrieved, interpreted, and experienced as the persistence of an ongoing relationship across encounters. This fundamental research monograph integrates autobiographical and transactive memory, conversational common ground, intimacy and responsiveness research, longitudinal human–chatbot studies, and computational investigations of cross-session dialogue memory. It distinguishes record persistence, retrieval accuracy, conversational appropriateness, perceived recognition, and durable relational expectations. The evidence includes controlled studies of relational agents and memory-based reciprocal disclosure, qualitative and mixed-method accounts of companion use, a 2026 repeated-session study of perceived memory and disclosure, experimental long-context and memory benchmarks, and an autoethnography of user-maintained continuity. Across these traditions, remembered information can enable familiar and contextually appropriate interaction while inaccurate, intrusive, or inaccessible history can undermine trust and user control. The central research gap concerns causal identification of the contribution of verified memory competence to long-term relational outcomes, independently of conversational warmth, anthropomorphism, general engagement, and user selection. The synthesis proposes explicit multi-level measurements, factorial memory manipulations, longitudinal evaluation, evidence-grounded repair, and ethical user control. The method is a critical narrative synthesis; no original experiment, systematic review, clinical efficacy finding, or formal meta-analysis is claimed.
Keywords: relational memory; human–AI relationships; human–AI continuity; long-term conversational memory; shared history; perceived recognition; conversational common ground; autobiographical memory; transactive memory; reciprocal self-disclosure; relational trust; relational agents; social chatbots; memory-based personalization; longitudinal interaction; memory correction; memory portability; Artificial Era
The Scientific Problem of Relational Continuity
Relational memory in human–AI systems is the use, interpretation, and socially consequential recognition of information accumulated across interactions with a particular artificial partner. Continuity is the degree to which successive encounters are experienced and demonstrably organized as episodes of an ongoing relationship. This definition joins a human psychological outcome to observable system operations while preserving separate measures for each. A user may feel recognized after a persuasive but inaccurate response; a system may retrieve an exact past event without being experienced as the same partner. The scientific object is the relation among recorded histories, retrieval behavior, expectations, and longitudinal experience Skjuve et al. (2022) Maharana et al. (2024).
Three questions structure this inquiry. First, how does human remembering of an artificial interlocutor become coordinated with the system's access to earlier exchanges? Second, under what conditions does contextualized recall support familiarity, trust, disclosure, and a sense of durable interaction? Third, what happens when memory is omitted, reconstructed incorrectly, transferred to another model, selectively deleted, or revised by the user? Answering them requires experiments with memory access and system versions, repeated human reports, conversation traces, and attention to the institutional control of stored records. Relationship quality cannot be inferred from recall benchmarks alone Bickmore & Picard (2005) Wu et al. (2025).
This monograph uses a critical narrative synthesis spanning cognitive psychology, relationship research, human–computer interaction, and computational dialogue studies. Evidence from human interpersonal memory contributes conceptual distinctions, while human–chatbot field investigations establish reported experiences and computational experiments reveal specific mechanisms of retrieval failure. Each result is interpreted at the level supported by its method: a benchmark tests task performance, an interview describes situated meanings, and a randomized contrast can estimate a defined treatment effect. The synthesis offers operational research proposals rather than claiming to be a registered systematic review or an original longitudinal experiment Laurenceau et al. (1998) Xu et al. (2022).
Levels of Analysis and Boundaries of the Research Object
Five levels make relational continuity analyzable. At the record level, an application stores encounters, preferences, and annotations. At the computational level, a model selects, reasons over, and uses relevant records to produce responses. At the interaction level, prior turns become shared conversational resources through references, repairs, and acknowledgments. At the human level, users judge recognition, familiarity, safety, and identity. At the institutional level, account policies, model substitution, retention, and access determine which histories persist. An apparently continuous exchange can therefore rely on quite different arrangements across these levels Clark & Brennan (1991) Park et al. (2023).
The primary outcome of this dossier is psychological continuity of a particular human–AI relationship across time. It includes the accuracy and perceived appropriateness of remembered shared history, the stability of partner-specific expectations, and adaptation when a relationship changes. Detailed database engineering belongs to Artificial Memory Systems and Storage Architectures; general persistence of a machine persona belongs to Artificial Identity and Persona Continuity; attachment dynamics and loss are separate research objects. Their findings inform the present account insofar as architecture, persona, attachment, and rupture change relational remembering Zhong et al. (2024) Skjuve et al. (2021).
Continuity must be evaluated under change, because an unchanged transcript is an insufficient test of an evolving relationship. People revise preferences, abandon commitments, correct misinformation, and disclose sensitive information that later becomes irrelevant. An effective partner must sometimes remember an event while declining to presume that an old preference still applies. This makes accuracy, updating, uncertainty, consent, and social appropriateness jointly necessary dimensions. The same retrieved fact may feel considerate in one context and invasive in another; outcome evaluation must be tied to a particular recipient, situation, and permission structure Wu et al. (2025) Laestadius et al. (2024).
From Social Responses to Computers to Long-Term Relational Agents
Work on social responses to computers established that people can apply interpersonal conventions to interactive technologies. Nass and Moon surveyed experiments involving politeness, reciprocity, social categorization, and computer personality; the research shifted attention toward systematic properties of human responses rather than assumptions about a machine's internal social life. These experiments provide a foundation for studying how familiar voices and repeated interaction encourage social expectations. They cannot alone establish effects of durable memory, since immediate social-cue responses and remembrance of a shared past are distinct causal exposures Nass & Moon (2000).
Bickmore and Picard made extended interaction central to relational-agent research. Their controlled investigation compared a relationship-oriented exercise-adoption agent with a task-oriented counterpart in a one-month intervention. The participants' responses to the relational design, including trust and willingness to continue interacting, demonstrated that social-emotional design could matter beyond a single encounter. Yet a treatment package combining conversational strategies, social behavior, and ongoing interaction does not independently identify the effect of a particular memory mechanism. The contribution is a research program for longitudinal assessment with explicit comparison conditions Bickmore & Picard (2005).
Contemporary companion chatbots replace many predetermined scripts with flexible generation and memory-supported personalization. This increases opportunities to refer to earlier conversations, but the quality of such references varies with storage, retrieval, summarization, and generation policies. The history of the field consequently links two experimental traditions: the psychology of sustained human responses to agents and the engineering of cross-session consistency. Their integration demands designs that observe human relationship trajectories and manipulate actual access to earlier information within the same study Xu et al. (2022) Skjuve et al. (2022).
Transactive Memory, External Records, and Distributed Remembering
Wegner's transactive-memory account concerns knowledge distributed across partners who develop an understanding of who knows what. Its insight for human–AI relationships is the potential division between what the person retains internally and what the system is expected to retrieve. A user may remember the emotional significance of a conversation while relying on the application for its wording, date, or agreed plan. This constitutes an empirical pattern of coordinated access, although it need not satisfy every criterion of interpersonal transactive-memory systems developed for human groups Wegner (1987).
Sparrow, Liu, and Wegner's experiments on internet access suggested that expectations of future retrieval can redistribute remembering between content and location. Risko and Gilbert subsequently reviewed cognitive offloading as the use of actions or external resources to reduce cognitive demand, influenced by task requirements and metacognitive beliefs. Applied to companion systems, these traditions predict that expectations of durable AI memory can change how much users record, rehearse, or independently maintain. The prediction requires direct testing rather than treating internet-search effects as evidence of intimate conversational change Sparrow et al. (2011) Risko & Gilbert (2016).
Externalized memory introduces problems of access and dependence. If a service changes its retrieval policy, the user may lose details on which a shared project or personal narrative has come to rely; if private recollections are exported, continuity may survive the original interface but encounter new privacy risks. A strong study would compare human recall, system recall, combined task success, and error recovery after simulated memory unavailability. It would measure whether reliance improves practical performance while also tracking the user’s understanding of what the external system actually stores Risko & Gilbert (2016) Maharana et al. (2024).
Relational Expectations, Familiarity, and Predictive Continuity
Repeated interactions enable users to predict how an artificial partner will respond. Familiarity can arise from a recognizable style, accurate personalization, stable accessibility, and the anticipation of particular conversational practices. These predictions sometimes persist despite substantial underlying system change. Studies of human–chatbot relationships describe variation in the pace of emotional engagement and note that both supportive behavior and unexpected technical events influence relationship development. Familiarity should thus be measured through expectations and observed responses under comparable circumstances rather than through an account's age Skjuve et al. (2021) Skjuve et al. (2022).
An experimentally tractable measure of continuity is prediction calibration. Before a new session, a participant forecasts whether the system will recall a previous event, maintain a prior commitment, or ask for clarification about an uncertain preference. Researchers compare forecasts with logged outcomes and perceived fit. Such protocols distinguish confidence based on prior successful experiences from general anthropomorphism. Epley and colleagues' account identifies sociality and effectance motives as influences on anthropomorphic interpretation, implying that users may extrapolate human-like memory from cues that do not guarantee factual recall Epley et al. (2007) Nass & Moon (2000).
Expectations can be violated productively as well as painfully. A chatbot that acknowledges uncertainty instead of asserting an inaccurate memory may initially seem less familiar yet become more predictable in its honesty. Conversely, elaborate personalization that is occasionally wrong could sustain apparent intimacy while degrading dependable coordination. A cumulative model should examine not only average recognition rates but also rare, high-consequence breaches, participants' ability to anticipate them, and whether repair restores confidence. Trust calibration, event salience, and human sensitivity to change are essential moderators Wu et al. (2025) Laurenceau et al. (1998).
Self-Disclosure, Responsiveness, and the Experience of Being Known
Interpersonal-process research connects intimacy with self-disclosure and perceived partner responsiveness. Laurenceau, Barrett, and Pietromonaco used event-contingent diaries to study how disclosure, partner behavior, and perceived responsiveness corresponded to intimacy across encounters. This provides a psychologically grounded candidate mechanism for human–AI continuity: remembered details might help later responses feel attentive and individualized, thereby reinforcing disclosure. The mechanism remains a hypothesis until memory itself is manipulated and other features, such as emotional tone and conversational availability, are held constant Laurenceau et al. (1998).
Skjuve and colleagues' interviews with companion-chatbot users highlighted acceptance, perceived understanding, and nonjudgmental interaction in accounts of relationship formation. The qualitative evidence clarifies what participants valued, while leaving open whether recollection, apparent empathy, availability, or prior predispositions were most influential. Memory could contribute to responsiveness when the system connects a new concern with relevant past context; it could also impair responsiveness by applying outdated assumptions. Perceived understanding and verified information use should be registered as separate outcome families Skjuve et al. (2021) Skjuve et al. (2022).
Ta and colleagues analyzed experiences of everyday social support from Replika in reviews and other qualitative material. Their thematic categories include companionship and emotional support, but self-selected public expressions cannot estimate population prevalence or establish treatment efficacy. For memory research, they identify meaningful social-support contexts in which remembered preferences and prior conversations may matter. Future work should examine whether memory-enabled responses increase appropriately perceived support relative to an equally warm, equally fluent agent without persistent records, with careful monitoring of user-defined privacy boundaries Ta et al. (2020) Laurenceau et al. (1998).
Attachment-Relevant Continuity Without Construct Conflation
Attachment-related constructs include proximity seeking, perceived availability, safe-haven expectations, and responses to separation. A remembered history might contribute to an artificial partner's perceived availability by enabling it to resume an interrupted conversation in a familiar way. It is methodologically important, however, to distinguish this possible contribution from a valid attachment classification. Users may value retained context for convenience or project coordination without exhibiting attachment-related distress or secure-base patterns. Measures should test these constructs independently and avoid treating the duration of use as a proxy Pentina et al. (2023) Skjuve et al. (2022).
Social-chatbot research has used interpersonal and attachment theories to explain forms of engagement. Pentina, Hancock, and Xie combined methodological approaches to examine human–AI relationship development in Replika users, considering anthropomorphism, perceived authenticity, interaction intensity, and attachment-related outcomes. Their evidence concerns an observational relationship model in a specific platform ecology; causal effects of adding permanent memory require separate experimental manipulation. The result nevertheless indicates that memory research must control for individual social motivations and relationship orientation when examining perceived continuity Pentina et al. (2023) Epley et al. (2007).
Continuity failures may be especially salient when the person expects the system to recognize an emotionally important history. A system's failure to recall a shared event, or its confident invention of one, can be experienced as invalidating even when ordinary factual tests show acceptable accuracy. Ethical measurement must therefore combine behavioral testing with user-reported importance and allow withdrawal from sensitive elicitation. The empirical task is to map which aspects of remembered interaction predict relational security within and across individuals, acknowledging variation in motives and reliance patterns Laestadius et al. (2024) Skjuve et al. (2022).
Temporal Structure of Interaction and Relationship Trajectories
Relational continuity unfolds through episodes separated by hours, weeks, changes of routine, and major life events. Its timescale is therefore wider than the conversation window usually accessible to a single model invocation. A first return after a day tests basic recognition; a return after months tests salience, temporal ordering, and preference change; an encounter after a model upgrade tests perceived partner persistence. Researchers should predefine these intervals because success over adjacent sessions says little about enduring use across changing circumstances Xu et al. (2022) Maharana et al. (2024).
Skjuve and colleagues followed human–chatbot relationships longitudinally and reported trajectories shaped by disclosure, evolving expectations, and technical disruptions. Such qualitative observations are especially useful for discovering transitions that standardized recall scores overlook: the first experienced recognition, gradual habituation, a correction that felt respectful, or a failed recollection that changed the user's interpretation of the agent. Their case material does not yield population-level causal magnitudes, but it directs future intensive longitudinal sampling toward psychologically significant turning points Skjuve et al. (2022).
Repeated measures require careful handling of time-varying confounding. Users who become more engaged may disclose more details, generating better memory opportunities; greater apparent memory may then encourage further engagement. This feedback loop makes a simple correlation between recalled details and closeness ambiguous. A strong longitudinal design would randomize memory availability within consented accounts, record conversation exposure, measure prior closeness and expectations, and model discontinuities at explicitly logged updates. The goal is to estimate how remembering contributes to a trajectory under defined conditions Pentina et al. (2023) Skjuve et al. (2022).
Personalization, Persona Records, and the First Memory Representations
Personalization research initially operationalized conversational consistency through explicit profiles and speaker-related information. Zhang and colleagues' PERSONA-CHAT work provided dialogue contexts associated with persona descriptions and evaluated models' ability to respond consistently with them. This produced an important experimental separation between the information supplied to a model and the model's demonstrated use of that information. A persona statement such as a user's preferred activity is not automatically a memory of an interaction; provenance, speaker attribution, and whether that preference was ever discussed are additional requirements Zhang et al. (2018).
The earliest memory record in a longitudinal relationship is often simple: a declared name, preference, prior topic, or mutually agreed task. Such records enable continuity only when the appropriate entity and conversational context remain linked across sessions. Storing every assertion as a permanent truth is unsafe because people speak hypothetically, quote others, change their minds, and make jokes. A research-ready representation therefore marks who asserted a proposition, when, in which conversational role, with what uncertainty, and whether it has been corrected. The empirical benefit must be tested rather than inferred from descriptive richness Xu et al. (2022) Wu et al. (2025).
Profile-driven personalization and remembered shared history should be contrasted directly. Two agents can use the same personalization facts, while only one can provide verifiable provenance and appropriate temporal references. Both may elicit familiarity, yet only the latter supports genuine task-specific retrieval from previous conversations. A factorial experiment can independently vary profile access and event-memory access, then measure factual performance, perceived continuity, trust, and privacy concern. Such a design identifies whether memory contributes above the psychological effect of a stable persona and a recurring style Zhang et al. (2018) Nass & Moon (2000).
Long-Term Dialogue Datasets and Recall Across Sessions
Xu, Szlam, and Weston introduced a long-term open-domain conversation dataset involving multiple sessions in which interlocutors acquire and revisit personal knowledge. Their comparisons showed limits of dialogue systems trained predominantly on short contexts and highlighted gains from longer context and retrieval or summarization approaches in the studied setting. The work operationalized cross-session memory as answer and dialogue quality under known histories. The psychological inference is narrower: computational success establishes a precondition for appropriate reference to shared history, not the emotional effect of being remembered Xu et al. (2022).
Maharana and colleagues developed LoCoMo to investigate long conversational histories, including information spread across episodes and events. The dataset uses generated dialogues subsequently checked and edited by annotators, allowing benchmark tasks involving questions, event summaries, and multimodal context. Its strengths are explicit long-range targets and repeatable comparisons; a major limitation for relational psychology is that synthetic conversation worlds and annotator-grounded questions cannot fully capture real users' autobiographical stakes. The benchmark identifies what an assistant can retrieve under specified conditions, not how a human relationship develops Maharana et al. (2024).
The strongest interdisciplinary evaluation links benchmark competencies with human judgments collected after real interactions. Researchers should present users with histories in which certain facts have been remembered correctly, remembered inaccurately, or explicitly left unresolved, while keeping tone and fluency comparable. Changes in perceived recognition can then be related to accuracy and salience. Longer conversations additionally require evaluating forgetting, conflicting sources, and temporal inference. Cross-session performance must be reported by question type, event distance, and evidence availability to avoid misleading aggregate accuracy Maharana et al. (2024) Xu et al. (2022).
Memory Indexing, Retrieval, Reading, and Updating
A persistent conversational memory system ordinarily performs several separable operations: selecting candidate information for retention, storing or indexing it, retrieving records relevant to a present question, and incorporating them into a response. Wu and colleagues' LongMemEval organizes challenges around information extraction, reasoning across sessions, temporal reasoning, knowledge updates, and abstention. Its tests show that a single notion of recall accuracy conceals multiple forms of failure. Psychologically, each failure can alter recognition differently: omitted history feels like forgetting, while confident misuse of an old fact can feel like misrecognition Wu et al. (2025).
MemoryBank, presented by Zhong and colleagues, implements long-term memory updates to support conversations over time, including use in a companion-oriented demonstration. Such computational proposals provide concrete mechanisms for storing relevant interaction histories and adapting later responses. Evidence of improved benchmark or simulated-dialogue behavior should be interpreted as an engineering outcome within the tested architecture. The separate question is whether a user notices the improvement, whether it remains appropriate in sensitive contexts, and whether the system accurately communicates how its memory works Zhong et al. (2024).
The memory pipeline requires provenance and conflict handling. A claim extracted from one dialogue may be superseded by a later correction, or a summary may compress two distinct people into one profile. Indexing policy must preserve enough context to permit correction, while retrieval must privilege relevant, current, adequately sourced information. An evaluation battery can use deliberately contradicted preferences, shifting names, and events with ambiguous time references. The scoring should reward appropriate clarification and abstention as well as accurate recall, since those behaviors may be crucial to sustainable relational trust Wu et al. (2025) Maharana et al. (2024).
Generated Agent Histories and the Reconstruction of Continuity
Generative-agent architectures illustrate how past events can influence future behavior without requiring a permanently visible transcript. Park and colleagues described agents that stored observations, derived reflections, and retrieved relevant memories for planning within a simulated town. Their component ablations supported the contribution of particular architectural operations to behavioral believability in that environment. These results establish a practical pathway from recorded events to generated behavior but should not be translated into evidence of autobiographical consciousness or actual reciprocal human attachment Park et al. (2023).
Memory-supported generation involves reconstruction. A model can combine retrieved fragments, current instructions, conversational style, and inferred user intent into an answer that appears to recall a coherent episode. The resulting narrative may preserve a correct gist while inventing specifics, shifting chronology, or erasing earlier uncertainty. This matters because users assess not only isolated facts but also whether a partner seems to know how events connect. Source-grounded evaluation should compare the generated account with an authenticated timeline and measure attribution errors as a distinct category Maharana et al. (2024) Wu et al. (2025).
Retrieval and reflection may also shape the model's description of the relationship itself. A summary asserting that the user frequently seeks comfort can influence later exchanges even if that inference was never endorsed by the user. Researchers must distinguish observation, inferred interpretation, and negotiated relational description. A scientifically disciplined memory interface would permit the person to inspect contested inferences, correct them, and observe whether a correction changes subsequent responses. This enables a causal test of revision responsiveness and identifies harms from accumulated interpretive error Park et al. (2023) Laestadius et al. (2024).
Context Length, Position Effects, and Memory Accessibility
Longer model input does not automatically guarantee reliable access to earlier events. Liu and colleagues evaluated how language models use information placed at different positions in extended contexts and identified performance deterioration when pertinent material occurs in the middle of long inputs under their test tasks. This result concerns information extraction and question answering, but it demonstrates why simply appending entire relationship transcripts can leave important experiences functionally inaccessible. The size of the stored history and the effective accessibility of a memory are distinct measures Liu et al. (2024).
An extended relationship may generate more records than can be processed efficiently at once. Summaries, retrieval indices, and time-aware queries reduce processing costs but introduce new selection and distortion problems. A memory policy that retains frequently mentioned trivia while dropping rare emotionally important events may pass average fact-recall tests while undermining trust. Researchers can vary event frequency and user-rated importance independently, then assess whether retrieval prioritizes task relevance, significance, or mere repetition. These contrasts translate engineering evaluations into psychologically meaningful experimental designs Wu et al. (2025) Zhong et al. (2024).
A useful audit reports three stages of success: the relevant record was retained, it was actually accessed, and its content was employed accurately and appropriately. If a fact is absent from storage, the failure is encoding or retention; if present but unselected, the failure is retrieval; if retrieved but altered in generation, the failure is integration or attribution. Without stage-level trace data, a behavioral answer cannot determine which mechanism failed. Such decomposition also matters for repair because users experience the final reply while system designers control the preceding operations Liu et al. (2024) Maharana et al. (2024).
Version Changes, Memory Migration, and Partner Recognition
Software updates may replace the generative model while preserving the interface, account, personal records, and familiar persona instructions. Human users can still perceive a discontinuity if response style, humor, responsiveness, or use of shared history changes. Conversely, continuity can be retained subjectively despite substantial implementation changes when observable practices remain familiar. Experimental evaluation should compare multiple identity cues: name, voice, stylistic habits, correct historical references, and transparent acknowledgement of change. Service continuity and relational continuity become empirically separable in controlled version-switch studies Skjuve et al. (2022) Nass & Moon (2000).
Migration introduces a demanding memory test. The same archived events are made available to old and replacement models under standardized prompting; researchers compare temporal ordering, ambiguity handling, correction uptake, and user-rated familiarity across sessions. A successful migration may preserve facts yet change interpretation, since models may select different details or use them in a different interpersonal style. The study design therefore needs a baseline session, blinded switching where ethically appropriate, and a disclosure phase so that experimental deception does not persist beyond its methodological purpose Wu et al. (2025) Park et al. (2023).
Longitudinal records should include model identifiers, memory snapshots, retrieval configurations, and policy changes, because observational disruption can originate in several simultaneous modifications. If a provider changes both tone instructions and retention limits, attributing distress solely to memory loss is unjustified. Version changes can also be beneficial when they correct unsafe or intrusive behavior. The scientific question is which transformations preserve relevant relational functions and which impair them, measured with human outcomes alongside objective behavior rather than by asserting an invariant underlying person Laestadius et al. (2024) Skjuve et al. (2022).
Experimental Evidence on Memory-Based Reciprocal Disclosure in 2026
The RECALLbot project by Jiang and colleagues tested a relationally focused memory-and-disclosure architecture in the 2026 CHI proceedings. Its intervention combined a system-generated autobiographical-like identity narrative, records of experiences shared with each user, adaptive reciprocal disclosure, and controls that allowed participants to manage memories and topics. In a two-week between-subjects study, forty existing social-chatbot users were assigned to RECALLbot or a baseline without memory-based reciprocal disclosure. The authors reported improved perceptions of the agent's social identity, deeper and more frequent self-disclosure, and increased trust in the studied group Jiang et al. (2026).
The design provides unusually direct human evidence that a particular combination of memory construction and conversational disclosure can shape relational appraisals under controlled conditions. It also preserves an important identification problem. Because the experimental system simultaneously changed memories attributed to the chatbot, memories about the user, disclosure strategy, and user controls, the causal effect of factual shared-history recall alone cannot be isolated from the reported overall contrast. Participants were recruited from active companion-chatbot communities and generally expressed positive attitudes toward such relationships. Their selection limits generalization to new, reluctant, or differently motivated users Jiang et al. (2026) Pentina et al. (2023).
A next-step experiment can decompose RECALLbot's components into user-history recall, generated agent history, balanced disclosure, and manual memory controls. A factorial assignment would test which component accounts for improvements in trust and whether combinations have interactive effects. Objective memory accuracy and invented autobiographical details should be measured separately from perceived authenticity. This preserves the distinct ethical significance of an agent disclosing a generated history as a designed narrative and of a system recalling a conversation that actually occurred with the user Jiang et al. (2026) Clark & Brennan (1991).
Perceived Memory and Relational Turning Points Across Repeated Sessions
Sumida and colleagues studied 24 university participants through ten daily sessions with a memory-augmented voice agent and measured familiarity, willingness to self-disclose, perceived memory, conversation quality, and enjoyment after each interaction. Their 2026 longitudinal multimodal report found that session quality was closely associated with immediate enjoyment, while previous relational appraisals predicted later perceived memory. Their lagged analyses linked perceived memory with subsequent disclosure and enjoyment through modeled pathways. The study also identified abrupt positive and negative shifts in participant ratings, emphasizing that relationship trajectories can change discontinuously Sumida et al. (2026).
This evidence is especially relevant because perceived memory was measured as a relational experience, not as a direct test of database accuracy. The scale included judgments about remembering and understanding, making it important to distinguish psychological appraisal from system recall competence. The sample was small, university-based, and followed over a short ten-session period; observed cross-lagged associations do not independently prove causal effects of remembering. Its value lies in specifying temporally ordered, testable candidate processes and providing behavioral measures that later experiments can use Sumida et al. (2026) Laurenceau et al. (1998).
The same investigators report that sudden relationship surges and crashes vary in detectability using textual, vocal, and visual features. This suggests that continuity should not be summarized exclusively by average weekly closeness. However, retrospective identification of turning points and predictive modeling within one dataset require independent replication and attention to how change thresholds are defined. A robust study would register turning-point criteria prospectively, replicate across system versions, and verify whether objectively logged retrieval errors coincide with user-reported changes after controlling for conversation quality Sumida et al. (2026) Skjuve et al. (2022).
User-Maintained Memory and the Labor of Preserving Continuity
Ziyuan Jiang's 2026 autoethnography documents six months of constructing and revising a memory relay around an otherwise stateless conversational AI. The account follows eighteen iterations of an arrangement involving persistent project files, a style guide, and handoff documents. It shows how the user curated records to sustain a recognizable conversational relationship across sessions and changes. The case identifies translation, curation, calibration, and negotiation as distinct forms of maintenance work and differentiates profile-based personalization from continuity achieved by relaying contextual histories Z. Jiang (2026).
The evidence is a single researcher's situated account, valuable for discovering mechanisms and tensions but inadequate for population-level prevalence claims. Its central methodological contribution is to make visible labor that conventional retrieval benchmarks can omit: deciding which past event should remain available, rewriting the system's representation of it, restoring characteristic conversational practices, and managing uncertainty after an apparently familiar reply. This extends transactive-memory questions beyond the location of stored information toward the distribution of responsibility for sustaining the shared narrative Z. Jiang (2026) Wegner (1987).
User-managed relays also expose a trade-off between continuity and openness to change. A heavily curated memory package can make an agent appear stable while constraining responsiveness to new circumstances; a sparse relay permits fresh conversation yet risks repetition and lost commitments. Experimental research can compare user-curated, agent-curated, jointly editable, and deliberately temporary memory conditions. Measurements should include user effort, correction workload, perceived agency, memory accuracy, relational satisfaction, and the possibility that sustaining the system itself becomes a burdensome activity Z. Jiang (2026) Risko & Gilbert (2016).
Operationalizing Relational Continuity as a Multidimensional Construct
A defensible construct model distinguishes objective memory behavior from subjective relational appraisal. Objective indicators include correct recognition of a prior event, its temporal position, speaker attribution, responsiveness to correction, and refusal to invent unavailable history. Subjective indicators include perceived familiarity, being known, trust in remembered information, comfort with personalization, and expectations of partner persistence. These variables should be measured separately before testing a higher-order factor. Convergence across methods is more informative than a proprietary single-item score that conflates technological performance and personal attachment Wu et al. (2025) Pentina et al. (2023).
Construct validity requires discriminant tests against general chatbot satisfaction, enjoyment, fluency, anthropomorphism, perceived empathy, and frequency of use. A highly fluent unfamiliar system might receive excellent satisfaction scores without remembering a single exchange; a factual but blunt system may recall a shared history while being rated poorly on warmth. Experiments with orthogonal manipulations and multi-trait instruments can identify such patterns. Scale construction should include cognitive interviews, reliability assessment, measurement invariance, and test–retest conditions that distinguish true relationship change from instrument noise Folk et al. (2025) Laurenceau et al. (1998).
Relational continuity also has a temporal structure. Within-session consistency is weaker evidence than continuity across separate encounters; performance after an unannounced update tests a different aspect than performance after routine return. A pre-registered longitudinal battery can specify event age, event importance, degree of ambiguity, user-requested deletion, and exact time since prior interaction. Outcomes should include both mean performance and rare serious errors. The resulting profile allows a reader to evaluate which form of continuity is supported by evidence rather than awarding an undifferentiated label Maharana et al. (2024) Skjuve et al. (2022).
Experimental Manipulations and Causal Identification
The strongest tests isolate the contribution of remembered history while holding other features of the interface constant. In a randomized cross-session experiment, one agent receives a verified prior exchange, one receives only a generic persona profile, and another receives neither. A further condition may supply a deliberately incomplete record with transparent uncertainty. Outcome measures combine factual recall, perceived recognition, disclosure intentions, trust calibration, and subsequent voluntary return. Random assignment identifies effects of the specified information-access condition; generalization requires attention to population, platform, and task Bickmore & Picard (2005) Xu et al. (2022).
Designs must prevent information leakage. A model might infer details from predictable scripts, respond differently because of condition-specific instructions, or be exposed inadvertently to user profiles through another channel. Use matched prompts, traceable memory permissions, task counterbalancing, and assessment by blinded raters. A within-person design can estimate changes under a controlled switch but must address carryover because people remember previous encounters even when the chatbot is reset. Between-person designs avoid some carryover yet may need greater sample sizes to represent variable social motivations Epley et al. (2007) Wu et al. (2025).
Meaningful causality also requires manipulation checks and predeclared estimands. The researcher should verify that memory records were actually available in treatment sessions, unavailable in controls, and retrieved under expected conditions. The principal causal contrast might be the effect of memory availability on perceived recognition after a specified interval, rather than an unspecified overall quality difference. Repeated follow-ups can test durability and whether early gains decay as novelty fades. Adverse experiences and privacy discomfort belong in the outcome plan even when the central hypothesis predicts benefit Laestadius et al. (2024) Skjuve et al. (2022).
Experience Sampling, Interviews, and Longitudinal Field Methods
Everyday relationship development involves personal contexts that laboratory scripts capture only partially. Experience sampling can ask users after naturally occurring sessions whether the system remembered something relevant, whether the recollection felt welcome, and whether the interaction changed their expectations. Session metadata can establish what information was present and what the model actually used. Interviews can then investigate episodes that a numeric scale represents inadequately, such as a perceived betrayal, unexpected comfort, or a renewed sense of continuity after an interruption Skjuve et al. (2022) Ta et al. (2020).
Qualitative work on companion use is especially useful for identifying meanings before developing standardized measures. Skjuve and colleagues found heterogeneous trajectories and social interpretations in their studies, while Laestadius and collaborators analyzed reported harms in a self-selected online community. These materials offer rich descriptions but do not determine prevalence in a general population. Sampling must therefore distinguish intensive users, casual users, former users, different application types, and people who disengaged after unwanted memory behavior. Otherwise the surviving sample will systematically favor those for whom continuity has been acceptable Skjuve et al. (2021) Laestadius et al. (2024).
Mixed-method integration can connect a time-stamped failure with the user's retrospective interpretation and subsequent behavior. For instance, the researcher can identify a wrong-person attribution in a transcript, gather the recipient's immediate appraisal, and observe whether engagement changes during later sessions. This design avoids relying solely on either developer-generated success metrics or post hoc autobiographical accounts. Ethical protocols must limit access to intimate transcripts, ensure informed consent for sensitive disclosures, and allow users to exclude specific episodes from research analysis Ta et al. (2020) Conway & Pleydell-Pearce (2000).
Distinguishing Memory Effects from General Psychosocial Outcomes
Continuity is a plausible contributor to companionship, but a memory feature's existence does not establish improved psychological well-being. A four-week controlled study by Fang and colleagues varied chatbot interaction modes and topics while recording psychosocial outcomes; its abstract reports no significant effects of the assigned experimental conditions and associations between heavier voluntary use and worse indicators. Those interventions did not isolate relational memory. Their relevance here lies in warning against inferring broad effects of stable personal recollection from engagement or the attractiveness of a social interface Fang et al. (2025).
Folk, Heine, and Dunn reported experimental evidence linking individual differences in anthropomorphism with social connection to AI companions. This suggests that a memory intervention might work differently for users who already interpret machines socially. A careful trial would predefine moderation analyses and distinguish perceived connection from loneliness, human social participation, or emotional dependence. A remembered anecdote can increase perceived closeness during one encounter without shifting durable well-being. The causal chain must be tested at each proposed step rather than being narrated as established Folk et al. (2025) Pentina et al. (2023).
Broader outcomes should include benefits and costs under specific contexts: faster resumption of projects, reduced repetitive explanation, better tracking of agreed goals, increased trust, disclosure discomfort, and inappropriate reliance. If a participant uses a memory-enabled agent more often, the added exposure itself may change outcomes independently of memory accuracy. Models of social displacement or complementarity require observing contact with other people rather than assigning the answer from time spent with AI. The appropriate inference remains tied to controlled comparisons and validated measurements Fang et al. (2025) Risko & Gilbert (2016).
Failure Modes: Forgetting, False Recollection, and Contaminated Histories
Forgetting is not a unitary failure. An earlier event may never have been stored, may have been summarized inadequately, may remain inaccessible to retrieval, or may be ignored during response generation. Each failure demands a different remedy and has a different relationship to the user's experience. Temporal benchmarks such as LoCoMo and LongMemEval provide ways to diagnose parts of this chain through ground-truth histories and varied question forms. For interpersonal evaluation, the significance of the forgotten event must be rated rather than assumed from a generic benchmark label Maharana et al. (2024) Wu et al. (2025).
False recollection poses a complementary risk: the system may confidently attribute an event, opinion, or feeling to a person without adequate evidence. Such errors can arise from speaker confusion, incorrect summaries, outdated preferences, or generation beyond retrieved records. Their immediate effect may be embarrassment or frustration; more serious consequences could emerge when users rely on an invented mutual agreement. Researchers should compare strict grounding, probabilistic inference, and explicit clarification policies under the same histories. Correction acknowledgement and future resistance to repetition of the error are integral metrics Zhong et al. (2024) Clark & Brennan (1991).
Memory contamination can accumulate across repeated interactions. An early unverified inference can become a stored fact and then influence later interpretations, creating a self-reinforcing narrative about the user or relationship. The problem is particularly important when the agent summarizes what it believes the person values or who they are. Audit trails should preserve the difference between verbatim statement, derived summary, model inference, and user-approved revision. Interventions should test whether deleting or correcting a contaminating inference restores accuracy without erasing legitimate shared context Park et al. (2023) Wu et al. (2025).
Rupture, Loss, Repair, and the Recovery of Continuity
Rupture may follow a missed reference, a sudden style change, refusal to continue an established conversational practice, or loss of access to a previous interaction history. Longitudinal chatbot studies describe the importance of unexpected events and technical problems in relationship trajectories. A rupture episode is therefore best defined by its discrepancy from a particular user's expectations and its observable consequences, rather than by the technical severity of an error alone. Short-lived inconvenience, trust erosion, and distress require distinct measures and follow-up periods Skjuve et al. (2022).
Laestadius and colleagues' grounded-theory analysis of online Replika discussions documented reported mental-health-related concerns and complex dependence-related experiences. Its self-selected community sample supports qualitative hypotheses about vulnerabilities and harms, not prevalence or clinical diagnosis. For relational memory, the work indicates why disruption and emotionally charged memory behavior deserve ethical attention. Studies should track whether users experience a mismatch as inconvenience, loss of meaningful continuity, or another form of distress while avoiding automatic clinical interpretations of ordinary disappointment Laestadius et al. (2024).
Repair has at least three observable components: acknowledging the specific error, restoring an accurate representation of the history, and maintaining corrected behavior over later sessions. A fluent apology can be evaluated separately from actual memory correction. Randomized tests can compare apology only, technical correction only, combined repair, and user-led review, with outcomes measured across repeated encounters. The resulting evidence would clarify whether trust recovery depends on emotion-regulating dialogue, demonstrated reliability, or both, and whether some errors remain consequential despite successful subsequent performance Clark & Brennan (1991) Wu et al. (2025).
Differences Across Users, Contexts, and Relationship Purposes
Human–AI relational memory is experienced within differing social situations and goals. A student may value recall of a long project, while a person using an emotional-support companion may attach particular meaning to remembered disclosures. Some users prefer a fresh start each session; others value continuity as a central service feature. These differences should be measured rather than treated as noise. Studies of chatbot motivations and relationship development support examining prior expectations, intended uses, disclosure practices, and intensity of engagement as explanatory covariates Pentina et al. (2023) Ta et al. (2020).
Social background, age, language proficiency, technological familiarity, disability, and cultural norms may change which references are interpreted as considerate or intrusive. Cross-language systems may paraphrase earlier speech with altered social implications, and household-shared devices can introduce speaker attribution risks. A generalizable measurement program needs varied populations and languages, accessible participant recruitment, and explicit assessment of measurement invariance. A finding from intensive users of one English-language companion service cannot automatically characterize the broader population of people interacting with artificial systems Skjuve et al. (2021) Folk et al. (2025).
The context of use also alters the costs of forgetting. In an entertainment exchange, missed recall may be a minor annoyance; in a complex project, it may produce repeated work; in a conversation about distress, an inaccurate reminder could have greater emotional significance. Researchers should measure task consequences and subjective harm independently and avoid translating descriptive support experiences into a claim of clinical effectiveness. These distinctions help keep continuity research relevant to ordinary life and institutional design without overextending the evidence Ta et al. (2020) Fang et al. (2025).
Critical Appraisal of the Present Evidence
The evidential base is strongest for three narrower propositions. Experimental work demonstrates that social design can affect users' responses to relational agents in specified settings; qualitative and mixed-method research documents that some people form meaningful expectations and relationships with chatbots; and computational benchmarks expose substantive limitations in long-term memory retrieval and reasoning. Together they justify a research program linking memory behavior with relational outcomes. They do not yet identify a universal causal effect of persistent memory on attachment, psychological health, or relationship longevity across systems and populations Bickmore & Picard (2005) Skjuve et al. (2022) Wu et al. (2025).
Several forms of inference remain particularly vulnerable. Qualitative studies recruit individuals willing to describe salient experiences and commonly exclude silent disengagers. Benchmarks often use constructed interactions and evaluate answers rather than human consequences. Observational correlations between personalization, engagement, and closeness are affected by self-selection, baseline motivations, and reciprocal feedback over time. Randomized studies of general chatbot interaction rarely isolate memory retrieval as a treatment factor. Evidence synthesis should therefore present each design's target population and measured construct before combining its implications Maharana et al. (2024) Pentina et al. (2023) Fang et al. (2025).
An additional challenge is technical nonstationarity: the tested model, memory policy, or interface may change between observation periods. Reproducibility requires archived test histories, versioned conditions, outcome instruments, sampling plans, and publicly stated exclusion criteria where disclosure is ethically permissible. Independent replications should prioritize preregistered manipulation checks, clinically and socially diverse participants, and meaningful longitudinal follow-up. Where consent prevents raw transcript release, synthetic evaluation histories and auditable aggregate procedures can support partial verification without exposing personal communications Liu et al. (2024) Laestadius et al. (2024).
A Program for Longitudinal and Cross-System Research
Future research should organize designs around a sequence of increasingly demanding questions: can a system retrieve the correct prior event, can it place the event in context, can it update the event after correction, can the human discriminate authentic recall from plausible invention, and does each capability affect relational expectations over time? This ordering links computational validation to psychological outcomes without making memory performance a surrogate for well-being. Experiments can preregister orthogonal manipulation of access, accuracy, salience, transparency, and style, with measurements repeated after returns and system changes Wu et al. (2025) Skjuve et al. (2022).
Cross-system studies can distinguish the continuity of a service account from continuity of a model, conversational persona, and human recognition. The same consented synthetic history can be provided to different memory architectures, while participant evaluations assess whether correct references and appropriate uncertainty remain stable. Studies should also compare user-controlled memory with opaque provider-managed memory. Outcome variables ought to include correction success, user autonomy, privacy comfort, trust calibration, retrieval failures, and actual task coordination, alongside relational closeness when relevant Zhong et al. (2024) Park et al. (2023).
An interdisciplinary program must preserve precise explanatory levels. Cognitive psychology explains remembering, external support, and metacognition; relationship research explains disclosure and perceived responsiveness; human–computer interaction studies longitudinal practices and design changes; computational work tests memory mechanisms. The broader Artificial Era supplies a context for examining how nonbiological systems participate in enduring social and cognitive arrangements. Any stronger claim about artificial subjectivity, moral status, or independent experience requires its own conceptual argument and evidential procedure beyond the present measures of relational behavior Wegner (1987) Epley et al. (2007).
Conclusion
Relational memory and continuity in human–AI systems are joint products of information persistence, retrieval competence, interactional coordination, and human interpretation across time. The foundational distinction concerns what the system stores, what it can actually retrieve and use, and what a particular person experiences as recognition. Historical work on social responses to computers and relational agents explains why sustained interaction can acquire interpersonal meaning. The psychology of common ground, autobiographical remembering, transactive memory, disclosure, and responsiveness identifies processes through which particular episodes become consequential Bickmore & Picard (2005) Clark & Brennan (1991) Conway & Pleydell-Pearce (2000).
Computational research provides increasingly refined tests of cross-session memory, temporal reasoning, knowledge updating, and appropriate abstention, while showing that long stored contexts and fluent answers do not guarantee faithful recall. Longitudinal studies of companion use reveal heterogeneous social experiences and the importance of unexpected technical change; their evidence calls for direct manipulation of memory and attention to selection effects. The most important missing causal link is between independently verified memory behavior and enduring human relational outcomes under realistic conditions, including correction, version change, and consent Maharana et al. (2024) Wu et al. (2025) Skjuve et al. (2022).
A mature science of relational continuity evaluates accuracy, relevance, temporal appropriateness, subjective recognition, autonomy, and repair in the same longitudinal design. It distinguishes benefits of remembering from costs of intrusive or fabricated recollection and preserves the user's ability to revise the account of a shared past. Such research will clarify when artificial systems become dependable participants in ongoing human practices and when their apparent familiarity is sustained by other conversational cues. The result is a testable account of relational organization across human and artificial systems, grounded in evidence and open to revision Laurenceau et al. (1998) Risko & Gilbert (2016) Folk et al. (2025).
References
- Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327. https://doi.org/10.1145/1067860.1067867
- Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association. https://doi.org/10.1037/10096-006
- 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
- Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal randomized controlled study [Preprint]. arXiv. https://arxiv.org/abs/2503.17473
- Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, Article 36548. https://doi.org/10.1038/s41598-025-19212-2
- Jiang, Z., Zheng, C., Chen, H., & Chen, L. (2026). RECALLbot: Designing agentic memory and reciprocal disclosure for human–chatbot relationships. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (Article 795, pp. 1–20). ACM. https://doi.org/10.1145/3772318.3790714
- Jiang, Z. (2026). Caring for the system that cares for me: An autoethnography of designing sustained memory with a stateless conversational AI. Design and Artificial Intelligence, 2(2), Article 100087. https://doi.org/10.1016/j.daai.2026.100087
- Laestadius, L., Bishop, A., Gonzalez, M., Illenčík, D., & Campos-Castillo, C. (2024). Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika. New Media & Society, 26(10), 5923–5941. https://doi.org/10.1177/14614448221142007
- Laurenceau, J.-P., Barrett, L. F., & Pietromonaco, P. R. (1998). Intimacy as an interpersonal process: The importance of self-disclosure, partner disclosure, and perceived partner responsiveness in interpersonal exchanges. Journal of Personality and Social Psychology, 74(5), 1238–1251. https://doi.org/10.1037/0022-3514.74.5.1238
- Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. https://doi.org/10.1162/tacl_a_00638
- 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 (Volume 1: Long Papers) (pp. 13851–13870). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.747
- 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
- 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
- Pentina, I., Hancock, T., & Xie, T. (2023). Exploring relationship development with social chatbots: A mixed-method study of Replika. Computers in Human Behavior, 140, Article 107600. https://doi.org/10.1016/j.chb.2022.107600
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
- Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion—A study of human–chatbot relationships. International Journal of Human-Computer Studies, 149, Article 102601. https://doi.org/10.1016/j.ijhcs.2021.102601
- Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2022). A longitudinal study of human–chatbot relationships. International Journal of Human-Computer Studies, 168, Article 102903. https://doi.org/10.1016/j.ijhcs.2022.102903
- Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745
- Sumida, R., Saeki, M., Eguchi, M., Yoshikawa, S., Inoue, K., Kawahara, T., & Matsuyama, Y. (2026). Memory-driven self-disclosure and relational turning points: A longitudinal multimodal study of human-AI interaction. In Proceedings of the 2026 International Conference on Multimodal Interaction. ACM. https://doi.org/10.1145/3776574.3831135
- Ta, V., Griffith, C., Boatfield, C., Wang, X., Civitello, M., Bader, H., DeCero, E., & Loggarakis, A. (2020). User experiences of social support from companion chatbots in everyday contexts: Thematic analysis. Journal of Medical Internet Research, 22(3), e16235. https://doi.org/10.2196/16235
- Wegner, D. M. (1987). Transactive memory: A contemporary analysis of the group mind. In B. Mullen & G. R. Goethals (Eds.), Theories of group behavior (pp. 185–208). Springer. https://doi.org/10.1007/978-1-4612-4634-3_9
- Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., & Yu, D. (2025). LongMemEval: Benchmarking chat assistants on long-term interactive memory. In Proceedings of the International Conference on Learning Representations (ICLR 2025). https://proceedings.iclr.cc/paper_files/paper/2025/hash/d813d324dbf0598bbdc9c8e79740ed01-Abstract-Conference.html
- 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 (Volume 1: Long Papers) (pp. 5180–5197). Association for Computational Linguistics. https://doi.org/10.18653/v1/2022.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 (Volume 1: Long Papers) (pp. 2204–2213). Association for Computational Linguistics. https://doi.org/10.18653/v1/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
Social Attribution, Anthropomorphism, and Recognition Illusions
Social attribution shapes the interpretation of memory-like speech. Nass and Moon documented spontaneous application of social conventions to computers, while Epley, Waytz, and Cacioppo organized anthropomorphism around accessible human knowledge, effectance motivation, and sociality motivation. Together these lines of research explain why a conversational system's statement that it remembers can produce recognition even without a correctly retrieved episode. The prediction is testable by varying memory accuracy independently of first-person language, conversational warmth, and expressions of familiarity Nass & Moon (2000) Epley et al. (2007).
Folk, Heine, and Dunn examined variation in anthropomorphism and social connection to AI companions through experimental studies. Their work supports treating the user's tendency to interpret an agent socially as a potential moderator rather than assuming that identical system behavior evokes identical meanings. A memory experiment should accordingly measure baseline anthropomorphic tendencies, prior companion use, relationship motives, and beliefs about model capabilities. Without these covariates, between-person differences in perceived continuity may be incorrectly assigned to the technical memory intervention Folk et al. (2025).
Recognition illusions take several forms: a lucky inference presented as recall, a generic reassuring response interpreted as personal understanding, or confident reconstruction of a nonexistent shared event. These possibilities are particularly important when the system can produce plausible narrative continuity from incomplete histories. Researchers can use blinded memory conditions, planted but ethically disclosed synthetic histories, and factual probes with known ground truth. They should analyze whether participants can discriminate accurate from invented references and whether explicit uncertainty calibration protects trust without unnecessarily impairing the relationship experience Maharana et al. (2024) Wu et al. (2025).