Remini: Chatbot for Mutual Reminiscence

• Remini is a chatbot that facilitates mutual reminiscence by guiding reciprocal sharing of positive memories between close partners.
• It employs a structured five-phase design—rapport, memory narration, elaboration, reflection, and summary—to scaffold emotionally rich and detailed dialogue.
• Empirical findings indicate that Remini significantly enhances positive affect, relationship quality, and narrative vividness compared to baseline chatbots.

Remini is a chatbot for mutual reminiscence: a technology-mediated process in which two people who experienced an event together revisit that shared positive memory through reciprocal self-disclosure, joint elaboration, and reflective dialogue. It was introduced as a system for close partners such as couples, friends, or family members, and is grounded in the Social Functions of Autobiographical Memory (SFAM) framework. In contrast to reminiscence systems centered on individual reflection, cue-based recall, or one-way storytelling, Remini is designed to scaffold emotionally rich, reciprocal remembering in a polyadic chat with two human participants and one chatbot (Jiang et al., 5 Aug 2025).

1. Conceptual scope and research problem

The motivating premise of Remini is that reminiscence is not only a private cognitive activity but also a social and relational practice. The paper defines mutual reminiscence as revisiting shared positive memories through reciprocal self-disclosure, and positions it as a mechanism that can strengthen emotional bonds, enhance well-being, and deepen intimacy. This orientation is important because many prior technology-mediated reminiscence tools are described as emphasizing either individual reminiscence, in which one person reflects alone, or social reminiscence, in which one person narrates to others in a narrator–audience format (Jiang et al., 5 Aug 2025).

The problem formulation is correspondingly narrow and specific. Prior tools often provide memory triggers such as photos, artifacts, sounds, or metadata, but do little to scaffold the conversational exchange itself. The identified gap is therefore not merely memory retrieval. It is the absence of support for reciprocal self-disclosure, emotional elaboration, partner responsiveness, and joint meaning-making. Remini addresses this by treating remembering together as an interactional process rather than a cue-response task.

This distinction is central to the system’s conceptual contribution. Mutual reminiscence, as framed here, depends on both partners contributing, responding, and co-constructing narrative detail. A plausible implication is that the relevant design problem in conversational systems is less about eliciting isolated recollections than about orchestrating a structured exchange in which memory narration, emotional disclosure, and relational reflection can emerge in sequence.

2. Theoretical grounding and system architecture

Remini is grounded in the Social Functions of Autobiographical Memory framework by Alea and Bluck, with three dimensions emphasized in the paper: memory narration characteristics, traits of memory sharers, and listener responsiveness. Memory narration characteristics refer to the vividness, emotional richness, and detail of a memory; traits of memory sharers include familiarity, closeness, age, gender, personality, and relationship length; listener responsiveness concerns active listening, empathy, and responsiveness as conditions for deeper self-disclosure and intimacy (Jiang et al., 5 Aug 2025).

These SFAM dimensions are translated into explicit design goals. The system is intended to support close, familiar relationships, encourage vivid and emotionally detailed narration, foster reciprocal participation, and make the listener or bot responsive and affirming. The result is not a generic memory-prompt generator but a guided conversational partner designed around the social function of remembering.

Technically, Remini is a Telegram-based chatbot powered by GPT-4o. It is implemented in Python, uses the Telegram Bot API, and uses GPT-4o (gpt-4o-2024-05-13) via the OpenAI API. The bot is embedded in a group chat with the two participants and is designed to look and behave like a regular group participant. It responds only when mentioned using @ReminiStory_Bot, and a Continue button allows participants to advance at their own pace. The system is also described as “ignorable,” so that human conversation remains primary.

The interaction is organized through a state-machine architecture with phase-specific prompts. Two LLM-based modules structure the workflow. The Conversation Driver generates the next message and decides when to move between phases. The Conversation Analyzer summarizes each completed phase and passes that context forward. The Driver receives the current phase chat history, phase-specific prompts, general prompts, and structured summaries from prior phases; the Analyzer produces a concise summary after each phase so that later prompts can be context-aware. This architecture makes continuity depend not only on raw dialogue history but also on structured intermediate representations.

3. Narrative pipeline and interaction logic

Remini’s core conversational design is a five-phase narrative pipeline: Rapport Building, Memory Narration, Elaboration, Reflection, and Summary. The phases move from initial social comfort to detailed recollection and then to reflective meaning-making, thereby structuring mutual reminiscence as a progression rather than as a single prompt sequence (Jiang et al., 5 Aug 2025).

In Rapport Building, the system greets users, asks how they are doing, invites introductions, shares a bit about itself to encourage reciprocity, asks about recreational activities and fun facts, and asks how the pair met and how the relationship has developed. The function of this phase is to create comfort and establish a friendly tone while eliciting basic reciprocal disclosure.

In Memory Narration, the chatbot asks whether users enjoy recalling happy memories, asks when they last discussed such memories, explains what counts as a positive memory, gives examples such as special trips, parties, and shared hobbies, asks the pair to choose one mutual memory, and then asks both users to write their memory concurrently and in detail. Participants are told to relax and clear their minds, write “as if it’s happening again,” include setting, environment, emotions, and events, and write more than 90 words. They are also told they can read each other’s narration afterward and continue when done. This phase converts simple cueing into explicit co-narration.

In Elaboration, the purpose is to enrich the memory with greater vividness. The chatbot asks which moment was most treasured, prompts users to deepen factual detail concerning people involved, places, objects, time, and context, and then asks about emotions. The paper explicitly links this stage to approaches for quantifying autobiographical memory content and to the idea that detail helps evoke richer remembering.

In Reflection, the system shifts from describing the event to interpreting it. Users are asked why the memory is treasured, how recalling it affects current emotions, whether the conversation makes them feel more emotionally connected, what new insights they gained about their relationship, and what they take away from the exchange. This is the phase in which emotional synchrony, relationship insight, connection, and deeper self-disclosure are most directly targeted.

In Summary, the chatbot consolidates the session by summarizing the reminiscence activity, including the narrations and follow-up answers, inviting gratitude and final remarks, and saying goodbye. The summary is intended to reinforce the meaning of the interaction and encourage appreciation between partners.

The prompt rules are correspondingly constrained. Remini is instructed to act as a reminiscence companion, do one task at a time, ask one question at a time, use simple, calm, respectful language, avoid bold text and bullet points, avoid terms like “phase” or “task” in user-facing language, always ask both users to answer together, actively listen by rephrasing user content, distinguish users by name or other considerate identifiers, and proceed only after receiving responses from both users. If a user does not respond, it may re-prompt; if both answer the last question in a phase, the Driver signals “phase done,” and the next phase begins.

4. Experimental design, participants, and measurement

The evaluation was a mixed-method, mixed-design experiment with a between-subjects factor of chatbot condition—Remini versus a baseline chatbot—and a within-subjects factor of phase or time—pre-interaction versus post-interaction. The sample comprised $N = 48$ participants organized as 24 dyads, including 15 male, 32 female, and 1 non-binary participant, aged 18–33 with median 22. Relationship types included lovers or significant others, friends, and siblings; all relationships had lasted over six months; all participants used Telegram regularly and were proficient in English (Jiang et al., 5 Aug 2025).

Participants were paired and matched across groups by age, gender, relationship type, relationship length, and personality traits. The Ten-Item Personality Inventory was used, and age, gender, relationship type, relationship length, and personality traits were included as covariates. The procedure was conducted remotely via Zoom from separate locations. Participants completed a pre-interaction survey, were assigned to either Remini or the baseline, entered a Telegram group chat with the experimenter and chatbot, and then interacted after the experimenter left. They could skip any chatbot question. After the conversation, they completed post surveys and interviews, and sessions ended with open-ended interviews. The study was IRB-approved under No. IRB-24-660.

Session duration differed substantially across conditions: baseline sessions lasted about 54 min ($SD = 15.5$), whereas Remini sessions lasted about 107 min ($SD = 47.7$). This difference is descriptively notable because the intervention was not merely a stylistic variation in prompting; it also changed the temporal extent of interaction.

Measurement combined chat-log coding and survey instruments. Quantitative chat-log measures were reminiscence duration, word count, message count, and words per message. Qualitative chat-log measures were number of details, diversity of details, and self-disclosure coding. Self-disclosure was coded in three categories—informational, thoughts, and feelings—with each rated from surface-level to deeply elaborated disclosure. Inter-rater reliability was reported as $\alpha = 0.803$ for detail presence or absence and Cohen’s $\kappa \ge 0.77$ for self-disclosure categories.

Survey measures included Positive Affect (PANAS positive subscale), Perceived Emotional Synchrony (shortened 6-item scale), Perceived Relationship Quality (shortened PRQC), Inclusion of Other in Self (single-item closeness measure), and Perceived Partner Responsiveness (4-item short form). Reported reliability included McDonald’s $\omega = .96$ for PRQ and McDonald’s $\omega = .91$ for PPR. Statistical analysis used Shapiro–Wilk to test normality, followed by ANCOVA, $t$-tests, ART ANOVA, Mann–Whitney $U$, Friedman, Wilcoxon Signed-Rank, and Dunn’s post-hoc tests with Bonferroni correction as appropriate.

5. Empirical findings

The principal quantitative result was that Positive Affect increased more in Remini than in the baseline condition. The ANCOVA interaction was $F_{1,37} = 6.33$, $SD = 15.5$0. In Remini, PA increased from 32.67 ($SD = 15.5$1) pre-interaction to 41.63 ($SD = 15.5$2) post-interaction; in the baseline, it increased from 31.04 ($SD = 15.5$3) to 35.46 ($SD = 15.5$4). The paper interprets this as a stronger positive mood boost under structured guidance (Jiang et al., 5 Aug 2025).

For relationship-related outcomes, Perceived Relationship Quality showed a main effect of chatbot design, $SD = 15.5$5, $SD = 15.5$6, and a main effect of phase, $SD = 15.5$7, $SD = 15.5$8. Median PRQ changed from 37 to 39 in Remini and from 35 to 36.5 in the baseline, with no interaction reported. Inclusion of Other in Self showed a main effect of phase, $SD = 15.5$9, $SD = 47.7$0, with median IOS increasing from 5 to 6 in Remini and from 4.5 to 5 in the baseline. The reported pattern is therefore not one of exclusive gains in only one condition; rather, both conditions increased closeness somewhat, while Remini was consistently higher.

Perceived Emotional Synchrony was higher in Remini than in the baseline, with $SD = 47.7$1, $SD = 47.7$2. Means were $SD = 47.7$3, $SD = 47.7$4 for Remini and $SD = 47.7$5, $SD = 47.7$6 for the baseline. Perceived Partner Responsiveness also favored Remini, with Mann–Whitney $SD = 47.7$7, $SD = 47.7$8; medians were 26.0 (IQR 24.0–28.0) for Remini and 24.0 (IQR 20.5–26.0) for the baseline.

Engagement differences were substantial. Remini produced significantly greater duration ($SD = 47.7$9, $\alpha = 0.803$0), message count ($\alpha = 0.803$1, $\alpha = 0.803$2), total words ($\alpha = 0.803$3, $\alpha = 0.803$4), and words per message ($\alpha = 0.803$5, $\alpha = 0.803$6). The medians were 57 versus 30 for duration, 63.5 versus 43 for messages, 891 versus 431.5 for total words, and 13.05 versus 6.7 for words per message.

Narrative vividness also differed sharply. Remini elicited a higher number of details ($\alpha = 0.803$7, $\alpha = 0.803$8) and greater diversity of details ($\alpha = 0.803$9, $\kappa \ge 0.77$0). Median number of details was 97.5 for Remini and 29.5 for the baseline; median diversity was 7.5 and 5.5, respectively.

Within the Remini condition, self-disclosure varied significantly across phases. Friedman tests were significant for informational disclosure, $\kappa \ge 0.77$1, $\kappa \ge 0.77$2; thoughts disclosure, $\kappa \ge 0.77$3, $\kappa \ge 0.77$4; and feelings disclosure, $\kappa \ge 0.77$5, $\kappa \ge 0.77$6. Informational medians were 1 in Phase 1, 3 in Phase 2, 3 in Phase 3, and 1 in Phase 4, with significant pairwise differences including P1 vs P2, P1 vs P3, P2 vs P4, and P3 vs P4. Thoughts disclosure medians were 1, 2, 1.5, and 2.5 across Phases 1–4; feelings disclosure medians were 1, 2, 2, and 2.5. The reported interpretation is that the phase structure shifted disclosure from basic factual sharing toward deeper reflective and emotional sharing.

6. Qualitative interpretation, limitations, and research significance

Interview findings complemented the quantitative results in three main themes. First, participants reported that the prompts facilitated in-depth reminiscence by helping them recall memories not discussed in a long time, bringing out forgotten details, helping them “relive” events, sustaining conversation when they might otherwise run out of things to say, and providing structure without entirely constraining the exchange. Some participants did report that the structure could be somewhat limiting. Second, participants described collaborative narration and reciprocal self-disclosure: each partner wrote an account, read the other’s version, filled in missing details, and discovered how the other interpreted the same event. Reported effects included a stronger sense of teamwork, mutual validation, seeing the memory “through the other’s eyes,” and wanting to create more memories together. Third, participants emphasized changed conversation dynamics. The chatbot was described as a therapist, a moderator, a facilitator, a mutual friend, and an observer. Some appreciated that it did not interrupt and could fade into the background; others wanted it to be more proactive; some found its questions algorithmic or the interaction more formal than ordinary conversation; and some were concerned about how it would interpret non-textual cues such as emojis or stickers (Jiang et al., 5 Aug 2025).

These findings support a nuanced view of chatbot mediation. The bot’s role was neither purely active nor purely passive. The design themes reported in the paper indicate that chatbot-mediated reminiscence works best when the bot is warm but not overbearing, helps users elaborate rather than dominating dialogue, encourages both partners equally, synthesizes and reflects what was said, is respectful of privacy and social boundaries, and adapts to a conversational rather than interrogative tone. The dual role of supportive mediator and near-invisible facilitator is therefore not incidental but constitutive of the design problem.

The paper identifies several limitations. The sample consisted of young adults from a single region, so cultural and linguistic norms may influence results. The comparison was only against a minimal-guidance chatbot, which means the study does not isolate the effect of chatbot presence from the effect of structured facilitation. The design was single-session, and relationship quality did not significantly improve in the short term, suggesting that longer-term or repeated use may be needed for stronger relational change. The authors also note potential anthropomorphism and privacy concerns, particularly in intimate group settings, and point to the need for richer modalities such as voice, emotion-aware responses, and more natural interaction styles.

In HCI and CSCW terms, Remini extends reminiscence research beyond personal reflection, cue-based memory retrieval, and narrator–audience storytelling by demonstrating technology support for mutual memory co-construction. In conversational AI terms, it shows that LLM-based chatbots can function as structured facilitators, reciprocal disclosure prompts, summarizers, emotional scaffolds, and conversation managers in a small group. The central practical lesson presented in the paper is that conversation design matters as much as model capability. A plausible implication is that the system’s observed benefits arise from the combination of phased interaction, turn-taking control, context-aware prompting, summary-based continuity, and empathy-oriented tone, rather than from the use of GPT-4o alone.