Echo Hui — PhD Enquiry

PhD Research Proposal · Prospective Enquiry

Knowing When Not to Recall

Memory Pacing for LLM Agents Supporting Scent-Linked Autobiographical Memory After Acquired Smell Loss

Echo Hui

In enquiry to the studio of Dr. RAY LC · City University of Hong Kong

Correspondence: hiechohui@gmail.com

1Abstract

Odour is a privileged retrieval cue for autobiographical memory: scent-evoked memories are older, more vivid, and more emotional than those evoked by other senses. For people with acquired smell loss, this pathway is severed — the memories remain encoded, but a class of cues that once opened them is gone. Long-term LLM memory agents could offer an alternative narrative pathway to these memories, yet current memory systems treat recall as a relevance-optimisation problem: they decide what to store and retrieve, not when a memory should surface. In sensitive autobiographical contexts the cost of mistiming is asymmetric — the same accurate memory can reconnect or re-wound depending on when and how it appears. This research proposes a two-stage inquiry. Stage 1 studies conversational, image-cued elicitation in which participants with acquired smell loss co-write their scent-linked memories with an agent that interviews, drafts, and revises under a grounding constraint — every written passage must be substantiated by what the participant shared, with the participant holding final editorial control. Stage 2 uses Wizard-of-Oz and scripted-prototype studies to compare memory disclosure policies — recall, cue, hold, ask consent, abstain — extending context-aware pacing from response timing in active listening to the timing and explicitness of memory disclosure. The work contributes memory pacing as a design problem, and names mistimed recall as a distinct, currently uncharacterised breakdown in sensitive human-AI dialogue.

2Problem Statement

When a privileged retrieval cue is permanently lost, access to a class of autobiographical memories is lost with it — not the memories themselves, but the door to them. For people with acquired smell loss, scent-linked memories of place, food, relationships, and childhood remain encoded but increasingly uncued.

A long-term conversational agent could serve as an alternative narrative pathway to these memories; but in this population, recall carries asymmetric cost. Surfacing a scent-linked memory can be a gift (reopening what was locked) or a wound (a reminder of what can no longer be smelled).

The core design problem is therefore not what an agent should remember, but when a memory should appear, how explicitly, and when the agent should hold, ask consent, or abstain — a question current memory systems, built around relevance optimisation, do not represent at all.

3Research Questions

This direction is intended as a starting point, to be refined in conversation with the supervising studio.

Stage 1 — Elicitation

RQ1

How do people with acquired smell loss describe their scent-linked autobiographical memories in relation to identity, place, relationships, and emotional continuity — and how has losing the olfactory cue changed when and whether these memories surface?

RQ2

Can conversational, image-cued co-writing serve as a substitute retrieval pathway for scent-linked memories, and what makes such elicitation feel like authorship rather than fabrication?

Stage 2 — Pacing

RQ3

What makes an agent's disclosure of a personal memory feel supportive versus intrusive, inaccurate, or emotionally unsafe — and how does this depend on timing, explicitness, and consent?

RQ4

How should an LLM memory agent decide among explicitly recalling, gently cueing, silently holding, asking consent, and abstaining over long-term interaction?

4Proposed Method: A Two-Stage Design

The research is structured as a two-stage pipeline in which the first stage supplies what the second stage requires: a memory disclosure policy can only be studied against a corpus of real, person-authored memories, and for people with acquired smell loss that corpus must first be made accessible without the olfactory cues that originally anchored it.

Stage 1

Elicitation

Image-cued conversational co-writing with people who have acquired smell loss. The agent interviews and drafts under a grounding constraint; the participant holds final editorial control.

produces a corpus that feeds

Stage 2

Pacing

Wizard-of-Oz comparison of disclosure policies: recall, cue, hold, ask consent, abstain.

Figure 1. The two-stage design. Stage 1 produces the person-authored memory corpus that Stage 2 requires in order to study when, and how explicitly, a memory should be disclosed.

Stage 1 — Elicitation

Co-constructing scent-linked memories through conversation and images.

Participants with acquired smell loss work with a conversational agent to elicit and co-write their scent-linked autobiographical memories, using participant-provided photographs of scent-associated objects, foods, and places as visual retrieval cues. The design premise is that when the privileged olfactory pathway to a memory is severed, visual and narrative cues can serve as a substitute retrieval pathway — reopening access to memories that remain intact but uncued. This inverts the direction of the studio's image-based memory work: rather than generating visuals to shape recall, participants' own photographs are used as cues to reopen it.

Methodologically, this stage adapts recent work on conversational autobiography writing (StorySage, Talaei et al., UIST 2025), adopting its central constraint of grounded co-writing: the agent may interview, draft, and revise narrative text, but every claim it writes must be substantiated by content the participant actually shared — no invented details, emotions, or elaborations — and the participant reviews and edits each draft, holding final editorial authority.

StorySage's own findings sharpen why this matters here: even with grounding rules, their system tended to overstate the emotional tone of memories, which participants experienced as inaccuracy. In a population where the accuracy and emotional register of a surfaced memory carry real weight, that unresolved failure mode becomes a first-class design and evaluation concern rather than a footnote.

Stage 1 outputs a person-authored memory corpus and a qualitative account of how these memories relate to identity, place, relationships, and emotional continuity (RQ1, RQ2). No olfactory hardware is involved at any stage.

Stage 2 — Pacing

When should the agent reopen the memory bank?

With each participant's own corpus in hand, a Wizard-of-Oz or scripted-prototype study compares memory disclosure policies — direct recall, gentle cueing, and a pacing-aware condition in which the agent may also hold, ask consent, or abstain. In the scripted-prototype variant, rule-based triggers govern disclosure timing while surface wording is template-constrained, with a researcher monitoring live for participant safety.

This extends context-aware pacing from response timing in active listening to the timing and explicitness of autobiographical memory disclosure, where the cost of mistiming is asymmetric: the same accurate memory can reconnect or re-wound depending on when and how it surfaces. Conditions are administered as staged, controlled conversational scenarios, following recent precedent for probing agent behaviour in emotionally sensitive dialogue.

Measures include felt understood, agency, trust, intrusiveness, emotional safety, and willingness to continue (RQ3, RQ4).

Stage 3 — Formalisation

Optional implementation extension — not a thesis contribution.

Findings from Stage 2 can inform a computational formulation of the disclosure decision (surface / cue / hold / consent / abstain) as a policy problem, evaluated on multi-session benchmarks. This is an engineering extension, not a thesis contribution.

The stages are deliberately ordered: Stage 1 establishes what the memories are and who owns them; Stage 2 establishes when an agent should touch them.

5Contribution

  1. C1

    Memory pacing as a design problem

    Extending context-aware pacing from active listening to the timing, explicitness, and consent structure of autobiographical memory disclosure in long-term agents.

  2. C2

    A named, currently uncharacterised failure mode

    Mistimed recall of an accurate, relevant memory is proposed as a distinct breakdown in sensitive human-AI dialogue — memory-timing-induced breakdown — extending existing breakdown taxonomies from response behaviour to memory behaviour.

  3. C3

    A population where the question has stakes

    Acquired smell loss makes disclosure timing consequential rather than cosmetic: the privileged olfactory cue is gone, the memories remain, and each surfacing can reconnect or re-wound.

6Connection to Studio Publications

Hear You in Silence

CHI 2026 · Honorable Mention→ Stage 2

Context-aware pacing for active listening — when an agent should respond, hold space, or stay silent. Stage 2 extends this from response pacing to memory disclosure pacing: when an agent should surface, cue, hold, ask consent, or abstain. This is the method spine of the proposal.

Tell Me What I Missed

CHI 2026→ Stage 1

Conversation as memory retrieval, studied for one-time witnessed events. Stage 1 extends this to long-term, scent-linked autobiographical memory built up across sessions, with the participant as author.

Regenerated Memories

ICHEC→ Stage 1

An image-based study of how visualisation reshapes remembering. Stage 1 works in the same modality but inverts it — using participants' own photographs as retrieval cues rather than generating visuals — and takes up its reconstruct-vs-distort tension as the grounded co-writing constraint.

Can AI Influence How We Remember?

ICHEC→ Stage 1

A companion study on how AI-generated visuals reshape recall. Together with Regenerated Memories it frames Stage 1's central design question: how to reopen a memory without distorting it.

7References

Studio Publications (Ray LC Lab)

  1. [1]

    Jiang, Chen, Zhang, Li & LC (2026). Hear You in Silence: Designing for Active Listening in Human Interaction with Conversational Agents Using Context-Aware Pacing. CHI 2026, Honorable Mention. DOI: 10.1145/3772318.3791238. Role: Method spine; Stage 2 extends response pacing to memory disclosure pacing.

  2. [2]

    Zhou, Gong, Shen & LC (2026). Tell Me What I Missed: Interacting with GPT during Recalling of One-Time Witnessed Events. CHI 2026. DOI: 10.1145/3772318.3791863. Role: Closest precedent for conversation-as-retrieval; Stage 1 extends it from one-time witnessed events to long-term, scent-linked autobiographical memory.

  3. [3]

    Wang, Xiao, Yeom, Gu & LC (2026). Regenerated Memories: How Remembrances can be Shaped by Image-based Visualization with Generative AI. ICHEC (Human-Engaged Computing). DOI: 10.1145/3786995.3787031. Role: Reconstruct-vs-distort tension; closest studio precedent for Stage 1's image-cued elicitation; motivates the grounded co-writing constraint.

  4. [4]

    Qian, Xiao, Liu & LC (2026). Can AI Influence How We Remember? Revisiting Memory Through AI-Generated Visuals. ICHEC (Human-Engaged Computing). DOI: 10.1145/3786995.3787032. Role: Same memory cluster; AI-generated visuals reshaping recall.

Non-Studio Publications

  1. [5]

    Talaei et al. (2025). StorySage: conversational autobiography writing, multi-agent. UIST 2025. DOI: 10.1145/3746059.3747681. Role: Stage 1 method precedent for grounded co-writing; their emotional-tone overstatement finding motivates emotional-register accuracy as an evaluation dimension.

  2. [6]

    Deng, Zhang, Jiao & Ren (2026). Breakdowns in Conversational AI: Interactional Failures in Emotionally and Ethically Sensitive Contexts. CHI 2026. DOI: 10.1145/3772318.3791186. Role: Stage 2 study-design precedent; the breakdown taxonomy that memory-timing-induced breakdown extends from response behaviour to memory behaviour.

  3. [7]

    Hackländer, Janssen & Bermeitinger (2019). An in-depth review of the methods, findings, and theories associated with odor-evoked autobiographical memory. Psychonomic Bulletin & Review, 26(2), 401–429. DOI: 10.3758/s13423-018-1545-3. Role: Grounds the notion of a privileged retrieval cue (Proust phenomenon).

  4. [8]

    Larsson, Willander, Karlsson & Arshamian (2014). Olfactory LOVER: behavioral and neural correlates of autobiographical odor memory. Frontiers in Psychology, 5:312. DOI: 10.3389/fpsyg.2014.00312. Role: Source of the LOVER characterisation — odour-evoked memories are Limbic, Old, Vivid, Emotional, and Rare.

  5. [9]

    Conway & Pleydell-Pearce (2000). The self-memory system. Psychological Review. Role: Memory as identity-construction, not storage; underpins narrative continuity.

  6. [10]

    Croy, Nordin & Hummel (2014). Olfactory disorders and quality of life. Chemical Senses. Role: Population motivation.

  7. [11]

    Tan et al. (2025). RMM: Reflective Memory Management; an RL reranker rewarded on citation usefulness. ACL 2025, pp. 8416–8439. Role: Gap evidence — even state-of-the-art memory management has no hold / consent / abstain in its action space; Stage 3 reference.

  8. [12]

    LongMemEval / LoCoMo. Multi-session memory benchmarks. Role: Stage 3 evaluation baseline.