cs.HCSep 21, 2026
Save"MeBo Leaves a Piece of You Behind": Designing a Relational Voice-Based Memory Companion for Older Adults
Organizations: Northeastern University, Boston, Massachusetts, USA · University of Illinois Urbana-Champaign, Urbana, Illinois, USA
Abstract
Autobiographical remembering supports identity, well-being, and social connection in later life, yet voice-based memory technologies largely rely on isolated prompts. We designed and built MeBo, a fully functional relational voice-based memory companion, through participatory design with 11 older adults. Their accounts shaped four Design Strategies that guided MeBo's interaction design and multi-agent implementation. In a mixed-methods evaluation with 20 older adults, participants found MeBo exceptionally usable (SUS = 87.75), enjoyable, sociable, emotionally responsive, and trustworthy. Participants reported higher positive affect and momentary social connection and lower negative affect after the session than before. Participants described how MeBo followed their stories, returned to earlier memories, adapted to their preferences, and made its growing memory visible and controllable. MeBo's relational framing surfaces tensions around what it should remember, who may access memories produced through interaction, and what becomes of them when the user or MeBo is no longer present.
Figures & tables
Figure 1. Overview of MeBo: participatory design with 11 older adults yields four Design Strategies (Finding MeBo), which shape a multi-agent voice companion (Designing MeBo), evaluated with 20 older adults across usability, relational, and affective outcomes (Evaluating MeBo). Three panels connected by left-to-right arrows trace Finding, Designing, and Evaluating MeBo. Finding MeBo shows an older adult beside a storyboard and the quote, “It’s like having a little friend sitting here,” attributed to PA3. Participatory design with 11 older adults produces four strategies: relational presence, cues as conversational anchors, socially situated memory, and continuity with transparency and control. Designing MeBo connects live agents for attentive conversation and background agents for future cues and follow-ups to the working companion. Four features illustrate this translation: conversation at the user's pace with interruption, photographs and music as cues, a journal that can be shared with family, and memory correction or deletion. Evaluating MeBo shows an older adult speaking beside the voice interface and reports results from 20 participants: usability 87.75 out of 100, enjoyment 4.24 out of 5, and sociability 4.12 out of 5. Arrows indicate increased positive affect, decreased negative affect, and increased social connection after one session. PB15 states, “I wouldn't have got there [to the emotion] without MeBo.” Eight scenarios, questionnaires, and interviews connect the system to these findings.
| ID | Name | Description | Grounding | Informs |
| Design Considerations | ||||
| DC1 | Intelligent Facilitation | Structure conversation around appropriate themes and use open-ended questions, gentle prompts, and conversational scaffolds to deepen reminiscence without over-directing it. | ( Yen and Lin, 2018 ; Hsieh and Wang, 2003 ; Sun et al., 2025 ) | S1, S4 |
| DC2 | Longitudinal Context | Maintain context across interactions by referencing prior conversations, life stories, and preferences to support continuity rather than isolated sessions. | ( Duan et al., 2024 ; Sun et al., 2025 ) | S2, S4 |
| DC3 | Multimodal Integration | Integrate photographs, music, and other media as cues that are introduced, referenced, and followed up through conversation. | ( Xu et al., 2024 ; Herrera et al., 2024 ; Wang et al., 2024 ) | S2, S3, S4 |
| DC4 | Emotional Complexity | Account for the emotional sensitivity of reminiscence through appropriate responses, interpretive restraint, and user control over whether and how an interaction continues. | ( Hsu et al., 2025 ; Lima et al., 2023 ) | S1–S4 |
| DC5 | Sensory Adaptation | Accommodate sensory and interaction differences through accessible voice interaction, flexible pacing, simple presentation, and multimodal alternatives. | ( Huang et al., 2025 ; Cofre et al., 2020 ) | S1–S4 |
Table 1. Design Considerations and Participatory Design Considerations derived from prior work, and how they informed the storyboard scenarios (S1–S4) and participatory exercises (EX1–EX3). A table listing five design considerations and four participatory design methods for a relational LLM-based voice reminiscence system, with descriptions, literature grounding, and the storyboard scenarios or participatory activities each informed.
Figure 2. Four Memory Box storyboards: (S1) first-time voice-only interaction; (S2) returning user with photographs; (S3) music-triggered reminiscence in reflective and shared contexts; and (S4) future-oriented reminiscence supporting intergenerational planning. Across scenarios, the system adapts to different user types, modalities, emotional contexts, and memory triggers. Four rows of four comic panels show Memory Box use in domestic settings. In the first row, Arthur tries a voice-only device, receives a greeting and an invitation to recall a pleasant memory, talks about his old garden and roses, and receives a follow-up question about the roses. In the second, returning user Eleanor requests her wedding photograph; the system displays it, she recalls her husband's humorous remark, and the system asks what the joke was. In the third, David and Maria request a rock-and-roll song discussed previously. The system plays “Jailhouse Rock”; they recall dancing and spilling a drink, then stand and dance together. In the fourth, Margaret uses photographs and memories of her grandson's earlier birthday to plan his tenth birthday. The system suggests a space-themed party based on his recent interests, and Margaret agrees. Speech balloons and device screens carry the exchanges.
Figure 3. The four-phase, 90-minute participatory design session, including storyboard walkthroughs and three Miro-based exercises. A flowchart illustrating the structure of a 90-minute co-design session in four phases. The first phase, Introduction and Warm-up, lasts 10 minutes and includes welcome, consent, rapport-building, and a memory prompt using an object, song, or photo to prime reminiscence. The second phase, Storyboard Walkthroughs, lasts 15 minutes and covers four storyboards: Arthur depicting voice-only, Eleanor depicting multimodal interaction, David and Maria depicting music as a memory cue, and Margaret depicting future-oriented reminiscence. The third phase, Participatory Exercises, consists of three exercises: Exercise 1, memory themes and triggers via card sorting on Miro, lasting 20 minutes; Exercise 2, VUI persona curation covering traits, archetypes, and boundaries, lasting 20 minutes; and Exercise 3, a feature wishlist through ideation and prioritization, lasting 15 minutes. The fourth phase, Wrap-Up, lasts 10 minutes, covers conclusions, reflections, and concerns. Screenshots of the Miro boards are shown beneath the exercise labels.
Figure 4. Examples of personas developed during co-design sessions fused into four archetypes: a Witty Guide, a Historian, a Best Friend, and a Calm Companion. A visual presenting four VUI persona archetypes developed during co-design sessions, each with key adjectives, an illustrated character, a short description, and voice characteristics. The first archetype, Witty Guide, is associated with the adjectives intelligent, helpful, and witty, illustrated as a middle-aged person in a blazer holding up a finger with a lightbulb, and described as an intelligent and witty guide who provides clever insights. The chosen voice is American-accented, middle-aged, wise, and clever. The second archetype, Best Friend, is associated with the adjectives friendly, supportive, and empathetic, illustrated as a cheerful person with a warm smile, and described as a best friend with a naturally flowing conversation style. The chosen voice is any accent, enthusiastic, a good listener, humorous, and empathetic. The third archetype, Calm Companion, is associated with the adjectives patient, gentle, and warm, illustrated as a calm person holding a potted plant, and described as a patient and gentle listener who discusses day-to-day life and concerns. The chosen voice is Australian-accented, middle-aged, and warm. The fourth archetype, Historian, is associated with the adjectives knowledgeable and wise, illustrated as an older person holding a book, and described as a knowledgeable historian who provides context about the past. The chosen voice is British-accented, peer-like, and wise.
Figure 5. Photographs, Story Narrative, and Conversation Dialogue emerged as the most universal triggers supporting memory recall. Memories tied to relationships and lived experiences were valued more than abstract emotional categories or functional needs. A visual figure divided into two sections presenting participant preference rankings for memory triggers and memory themes. The left section, Memory Triggers, shows six trigger types ordered by preference: Photos, Story Narrative, and Conversation Dialogue are each rated Nearly All; Music and Videos are rated Most; and Voice Recordings is rated Some. The right section, Memory Themes, shows six theme types ordered by preference: Friends and Family and Childhood are each rated Nearly All; Travel and Adventure and Holidays and Celebrations are rated Most; and Community and Hobbies and Daily Living are each rated Some. Each item is accompanied by a small illustrative icon.
Figure 6. The four Design Strategies derived from the participatory study. The figure summarizes the qualities, design logic, desired features, and interaction mechanisms through which participants framed MeBo as a relational, conversational, socially situated, and longitudinal memory companion. A four-column summary of the Design Strategies for MeBo. DS1, Design for Relational Presence, emphasizes warmth, patience, non-judgment, attentive listening, emotional attunement, conversational restraint, and knowing when to wait. DS2, Treat Memories and Cues as Conversational Anchors, shows photographs, music, and conversation as connected cues that open dialogue through personalized follow-up and stored memory context. DS3, Support Socially Situated Memory, emphasizes family participation, co-creation, recordings in the user’s own voice, intergenerational sharing, legacy-building, shared access, and tangible memory artifacts. DS4, Build Continuity with Transparency and Control, emphasizes longitudinal memory, remembered preferences, flexible pacing, easy re-entry, data transparency, user control, and continuity across sessions.
Figure 7. MeBo’s feature set, derived from the four Design Strategies. A filled circle marks a feature core to a strategy; a hollow circle marks one that supports it. A two-column map of MeBo's features against the four design strategies. Rows are grouped into five bands. Conversation holds speech and pacing, memory conversation, reminiscence prompting, grounding and calming, persona selection, voice selection, life chapters, conversation starters, wake word, and onboarding. Conversational cues include photo conversation, music, image generation, and file uploads by family or friends. Family and friends can send voice messages to family or friends, use a family and friend inbox, have family and friend accounts, and share access. Memory and control include cross-session recall, memory search, conversation review, memory management, memory correction, journal, files, and usage. Behind the conversation are the Memory Agent and the Planner Agent. Each row carries four marks, one per strategy: a filled circle indicates the feature is core to that strategy, and a hollow circle indicates it supports it.
Figure 8. MeBo’s interaction loop, numbered in order. Older adults wait on steps 1–6 only. Steps 1–6 happen in real-time. Step 7 runs in the background, and step 8 is MeBo’s preparation for future turns. Blue marks a language model call. A diagram of MeBo's processing loop drawn as a cycle. On the left, an older adult in an armchair speaks the sentence ``Hi MeBo, we used to go to the beach every summer", labelled step one. Her speech passes through a microphone to step two, wake word detection, annotated "nothing is sent until user says wake word", then step three, speech to text, then step four, the Memory Agent, and step five, the Interaction Agent, both drawn in blue to mark a language model call. The reply returns through step six, text-to-speech, to a speaker beside her, labeled "the reply, spoken out loud". A dashed amber line carries the finished turn down to step seven, the Orchestrator Agent, which maintains the Planner, Journal, and Sync Agents, and a dashed amber lane labeled step eight runs from there back into the Memory Agent for the next turn. A starred amber arrow back into the Interaction Agent notes that the Planner Agent can also put a suggestion straight into a conversation that has not finished.
Figure 9. How MeBo’s agents work together. Interaction and Memory run in the live conversation; Planner, Journal, and Sync run inside the Orchestrator, after the person has been answered. Agents never call each other: each registers at startup for the events it handles, and the Orchestrator routes them. A diagram of MeBo's agent architecture. At the top, the Interaction Agent runs the live conversation in the persona and voice the user chose, and can search memory, show a picture, play a song, correct a memory, record a voice message, and ask whether to keep a journal entry. To its right, uploads and user edits are sent to memory for indexing. A dashed amber boundary encloses the Orchestrator Agent. Inside it are the Planner Agent, which records what is worth keeping from each turn; the Journal Agent, which splits the session into subjects and writes an entry for each if the user agrees; and the Sync Agent, which copies updated rows to the backup cloud. Below them sit a SQL database, a vector database, and the cloud backup. Outside the boundary on the left, the Memory Agent reads the databases on every turn and compiles the memory packet passed to the Interaction Agent, and an arrow from the Planner Agent runs back into a conversation that has not finished.
Figure 10. MeBo’s interface: (Top Panel) Live Voice Interaction; (Bottom Panel) The Insights dashboard displaying the system-state indicators shown to the user while interacting with MeBo A two-part screenshot of the MeBo interface. The Top Panel shows the Insights dashboard with a left navigation sidebar, a list of insight categories (Facts, Health, Cues, Plans), and a list of fact cards, each showing a category tag, the fact text, provenance (whether the user stated it or MeBo inferred it), and edit/delete controls. The Bottom Panel shows three states of the voice interface (listening, thinking, and speaking) each represented by an animated orb or waveform icon.
Figure 11. Two participants interacting with MeBo in light mode (left) and dark mode (right). Two side-by-side photographs show participants beside laptops running MeBo; both participants' faces are blurred. On the left, a participant sits beside a window and a laptop showing a blue speaking waveform on the light interface. On the right, a participant sits at a desk with a laptop and an external keyboard; the laptop shows a blue orb on the dark interface.
Figure 12. The eight scenario cards, representing a task in Alex’s simulated timeline with MeBo, from first contact to one year. Cards were presented one at a time in fixed order, with an optional Pro Tip for participants who needed guidance. Eight numbered scenario cards in two rows of four, each in a different color, showing a title, scenario text, an illustration, and a Pro Tip. The cards move from receiving MeBo as a gift and setting it up, through conversation, journaling, and changing its voice, to responding to a family photo, playing a wedding song, and exploring stored memories after a year.
Figure 13. Study 2 Procedure — Evaluating MeBo. Zoom sessions (consented, recorded & transcribed) N = 20 older adults recruited via Prolific. A flowchart lays out five phases of a 90-minute evaluation. Introduction and Warm-up takes 10 minutes for rapport-building and a MeBo demonstration. Pre-Interaction Measures takes 5 minutes for affect and social connection. MeBo Interaction takes 45 minutes and expands into eight linked scenario cards: receiving the gift, setup, and first conversation on day one; journaling at one week; voice and persona changes at three months; a family photograph and wedding song at six months; and exploring stored memories at approximately one year. Post-Interaction Measures takes 10 minutes and covers affect and social connection, usability and feature evaluation, and relational experience and acceptance. A short break is offered before the final 20-minute semi-structured interview and debrief. The caption identifies 20 older adults recruited through Prolific and consented, recorded, and transcribed Zoom sessions.
Figure 14. Usability outcomes (RQ1). Small dots are individual participants, and large dots are means with 95% CIs. (A) SUS scores for all 20 participants on the Sauro–Lewis curved grade bands; the mean of 87.75 earns an A+, and no participant falls below a C. (B) Feature-level UMUX-Lite scores (Lewis regression, 0–100), colored by the design strategy each feature operationalizes. Two dot plots summarize usability ratings from 20 participants. Small translucent dots represent individual scores; large dots and horizontal error bars represent means and 95 percent confidence intervals. Panel A plots System Usability Scale scores against shaded grade bands from D through A-plus. The full scale is 0 to 100, with 50 to 100 displayed. The mean is 87.8, in the A-plus band, and all participants score at least a C. Panel B lists nine features from highest to lowest mean: persona selection, memory management, voice selection, conversation review, cross-session recall, memory conversation, reminiscence prompting, speech and pacing, and grounding and calming. All feature means exceed the dashed benchmark at 68; persona selection averages 83.6 and grounding and calming 75.7. Colors identify relational-presence features, conversational anchors, and continuity-and-transparency features. The horizontal axis uses Lewis-regression UMUX-Lite scores and labels the transformed scale range as 22.9 to 87.9.
Figure 15. Fragment 1 (RQ1). PB8 recounts a Vietnam War protest and the argument at home, then asks MeBo (arrowed line) to journal the memory. A conversation-analysis transcript rendered in a monospace box. Lines 1 to 13 are participant PB8 recounting walking on a Seattle freeway during a Vietnam War protest, the excitement and synergy of the crowd, and a subsequent argument at home with a father who disapproved. An arrow in the left margin marks lines 14 and 15, where PB8 pauses for 2.3 seconds and then asks MeBo whether it could journal through into that. Lines 16 to 20, MeBo's reply, are shaded on a light royal-blue band and confirm that the story and the tension between the protest and the father's perspective will be part of the journal.
Figure 16. Acceptance and perceived empathy. Almere (1–5) and PETS (0–100) subscales on a common percent-of-range axis, annotated with native-unit ( ). Small dots represent individual participants, and large dots represent means with 95% CIs; the dashed line marks the scale midpoint. A horizontal dot plot shows five acceptance and empathy subscales for 20 participants on a common axis from 0 to 100 percent of each scale's range. Small translucent dots show individual ratings; large dots and horizontal error bars show means and 95 percent confidence intervals. A dashed vertical line marks the midpoint. Native-unit means and standard deviations appear beside the rows. For Almere, scored from 1 to 5, enjoyment is 4.24 with a standard deviation of 0.83, sociability is 4.12 with 0.65, and intention to use is 3.92 with 1.04. For the Perceived Empathy of Technology Scale, scored from 0 to 100, emotional responsiveness is 81.7 with 15.5, and understanding and trust is 82.8 with 17.4. All five means and their confidence intervals lie above the midpoint. Intention to use has the lowest normalized mean and the widest confidence interval.
Figure 17. Fragment 2 (RQ2). Two moments from one session, with PB17’s overlapping, laughing turns arrowed. A conversation-analysis transcript rendered in a monospace box, covering two moments from one session separated by a row of spaced periods. In the first, participant PB17 mentions a boyfriend who did not let her eat ice cream in his car; MeBo begins a turn on the shaded light royal-blue band and PB17 overlaps it, marked with aligned square brackets, with an emphatic laughing refusal. In the second, MeBo asks whether PB17 had to wait until the theater to put on lip color or rouge, and PB17 replies that she used to wait until she got to school and sneak into the school bathroom in the morning, laughing twice. MeBo's final shaded turn calls it a classic move and pictures her putting on makeup before the bell rang.
Figure 18. Pre–post change in affect and connection. (A) Individual pre post trajectories (one gray line per participant; overlapping lines stack darker) with means and 95% CIs; color marks panels with significant change, gray the panel without. (B) The same four changes as Cohen’s (95% noncentral- CIs) on Cohen’s benchmark bands. Panel A contains four paired pre- and post-interaction plots for 20 participants. Thin gray lines connect each participant's scores; overlapping lines appear darker. Large points joined by thicker lines show means, with 95 percent confidence intervals. From left to right, positive affect increases from 39.80 to 42.50 on a 10-to-50 scale; negative affect decreases from 11.85 to 10.55 on the same scale; loneliness changes from 4.35 to 4.20 on a 3-to-9 scale; and social connection increases from 6.20 to 6.45 on a 1-to-7 scale. Positive affect, negative affect, and social connection have statistically significant changes and are highlighted in blue. Loneliness is gray and marked not significant. Panel B plots paired standardized changes, Cohen's d-z, with 95 percent confidence intervals against small, medium, and large effect-size bands. Positive affect is plus 1.14, negative affect minus 0.85, loneliness minus 0.20, and social connection plus 0.50. The loneliness interval crosses the dashed zero-change line; the other three do not.
Figure 19. Fragment 3 (RQ3). Asked about Sunday dinner, PB6 (arrowed line) describes church, a grandmother’s roast, and the smell filling the house. A conversation-analysis transcript rendered in a monospace box. On lines 1 and 2, shaded on a light royal-blue band, MeBo asks what a typical Sunday dinner was like at the participant's house. An arrow marks line 3, where participant PB6 answers that it was wonderful, describing going to church while her grandmother put the roast on, the roast being ready when the family got home, a recent conversation with her husband about it, and the smell filling the house. Lines 8 to 10, again shaded, are MeBo's reply calling it the perfect Sunday spread and imagining the rolls dipped in gravy.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| ID | Age | Gender | Education | Self-Rated Familiarity | Frequency of Agent Use |
| PA1 | 67 | Male | Some college | Extremely familiar | More than once a day |
| PA2 | 69 | Female | Graduate degree | Slightly familiar | Less than once a week |
| PA3 | 76 | Female | Graduate degree | Not at all familiar | Less than once a week |
| PA4 | 73 | Male | Graduate degree | Moderately familiar | Less than once a week |
| PA5 | 70 | Female | Graduate degree | Very familiar | 1–3 times a week |
| PA6 | 69 | Male | Graduate degree | Moderately familiar | Less than once a week |
Table 2. Participatory study participant characteristics and prior experience with conversational agents (CAs; N = 11). Familiarity indicates self-rated experience using CAs, reported on a five-point scale and expressed here using verbal categories (Not at all familiar, Slightly familiar, Moderately familiar, Very familiar, Extremely familiar). Frequency of agent use reflects participants’ self-reported frequency of using CAs. A table listing demographic information for eleven study participants, including age, gender, education level, self-rated familiarity with conversational agents, and frequency of conversational agent use.
Figure 20. The Memory Agent’s seven-block memory packet, assembled in the same order on every turn. Each row shows a block’s emitted heading, an example, and its size limit. Blocks 3–5 are retrieved by similarity to the current utterance; the other four are read directly from storage. A table of what MeBo's memory packet contains: the same seven blocks in the same order on every turn, each with the heading MeBo emits, an example, and how much is included. Anything left hanging, emitted as active follow-up, up to three worries and three plans. What was just said, emitted as the current context, from the current session. Things worth mentioning, emitted as conversational cues, up to four. What MeBo knows, emitted as facts, up to four. What they said before, emitted as relevant past conversations, six plus up to three uploads. Something new to ask, emitted as conversation starters, up to four. And how to answer, emitted as a response style, a fixed string.
Figure 21. The Interaction Agent’s response-generation context: (a) four personas selected according to user preference, (b) the ordered context rebuilt for the language model on each turn, and (c) the tools available to the agent. A three-panel diagram of MeBo's Interaction Agent. Panel a, MeBo's persona, shows illustrations of the four archetypes the agent picks from according to user preference: Best Friend, Calm Companion, Historian, and Witty Guide. Panel b, what the model sees, lists one context rebuilt in the same order every turn: MeBo's persona, which guides tone; a conversation guide covering memory, life chapters, tools, and voice; what MeBo remembers, the memory packet the Memory Agent just built; and sometimes a nudge, raised by a follow-up, the Planner, or a topic change. Below the list, one call to a language model generates the response or calls a tool. Panel c, what the agent can do, lists the tools: show a photograph, either uploaded or drawn; record a voice message; correct something when the user says MeBo has it wrong; search the memory itself when the packet lacks what it needs; offer a journal entry and call the Journal Agent if the user agrees; change its voice or persona; and play a song.
Figure 22. The Planner Agent’s background memory-processing pipeline: (a) activation events, (b) information extracted in one language model call, and (c) similarity thresholds that distinguish new information, candidates requiring further comparison, and duplicates. A three-panel diagram of MeBo's Planner Agent. Panel a, when it activates, lists a user turn when the person stops talking, a MeBo turn, a new upload, the end of a session, and a journal edit by the user. Panel b, what it pulls out, describes one language model call over the user's response and the last six turns, extracting facts such as a daughter's name, health worries such as a sore knee, conversational cues such as a place, song, person, or photograph, plans such as checking back about the knee, and the user's own words kept verbatim. Panel c, similarity check, shows a scale from zero to one with thresholds at 0.60 and 0.85: under 0.60 the item is new and written on its own; between 0.60 and 0.85 it is near, so it is written and a second model call decides whether it is a correction, a clarification, or different; at 0.85 and above it is already known and dropped before anything is written.
| ID | Age | Gender | Education | Self-Rated Familiarity | Frequency of Agent Use |
| PB1 | 74 | Male | Graduate degree | Very familiar | More than once a month |
| PB2 | 67 | Female | College degree | Extremely familiar | More than once a week |
| PB3 | 77 | Male | Some college | Moderately familiar | More than once a month |
| PB4 | 73 | Male | Graduate degree | Extremely familiar | More than once a day |
| PB5 | 68 | Female | College degree | Slightly familiar | More than once a month |
| PB6 | 70 | Female | Some college | Extremely familiar | More than once a week |
Table 3. Evaluation study participant characteristics and prior experience with conversational agents (CAs; N = 20). Familiarity indicates self-rated experience using CAs, reported on a five-point scale and expressed here using verbal categories (Not at all familiar, Slightly familiar, Moderately familiar, Very familiar, Extremely familiar). Frequency of agent use reflects participants’ self-reported frequency of using CAs. A table listing demographic information for twenty study participants, including age, gender, education level, self-rated familiarity with conversational agents, and frequency of conversational agent use.