In The digital therapist? LLMs and the future of clinical dialogue, which I co-authored with Ran Pei and Lluís Barceló-Coblijn, we explored a question: when large language models enter clinical communication and mental health settings, can they become “digital therapists”?
Our answer is cautious. LLMs can generate fluent, gentle, apparently empathic language. They can also assist with documentation, information retrieval, structured reflection, and carefully scoped communication. These capabilities can contribute to care, but they do not establish that a system is a therapist. Our argument places interpretive responsibility, clinical judgment, and moral accountability with the people and institutions delivering care.
The questions raised in that paper also inform the work of the HuChuan team.
HuChuan (湖川) focuses on AI long-term memory technology. Its aim is to support interactions that develop over time, preserving relevant information about a person’s experiences, preferences, emotions, relationships, and changing circumstances. HuBan (湖伴) is the product exploration that brings this technology into everyday psychological companionship.
We describe HuBan as an AI companion. It is designed to support expression, reflection, and continuity across conversations, while maintaining clear boundaries with professional counseling and clinical judgment.
This work brings us to two further questions: can AI really remember you, and when can memory-supported interaction become meaningful companionship?
Remembering someone involves more than recalling an isolated detail. Consider a person who once described a difficult experience and later explains that their feelings about it have changed. A useful companion would need to recognize the relationship between these accounts, preserve their temporal context, and remain open to the person’s revised interpretation. The quality of memory therefore concerns how experiences remain connected and revisable.
Our autobiographical-memory-inspired architecture approaches this problem through several complementary structures. It organizes information into raw dialogue, contextual snapshots, narratives, and structured events. Memory chains connect related records over time, while memory belts organize retrieval by context or domain. The aim is to preserve the relationship between a compact statement and the experience from which it was derived.
At a basic level, our “ten dimensions, four axes” framework combines ten content categories with four management axes.
The ten categories cover context, facts, viewpoints, preferences, observations, experiences, temporal information, emotions, relationships, and growth. These categories provide an engineering vocabulary for organizing the different kinds of information that emerge during sustained interaction.
The four axes concern:
-
Time: when information applies and how it relates to earlier and later experiences.
-
Importance: its relevance and priority for subsequent interaction.
-
Evidence: the sources supporting a record, including the distinction between user statements and system interpretations.
-
Status: whether information remains current, has changed, or requires clarification.
Together, these structures are intended to support continuity while making the basis of a response easier to inspect. Their contribution to companionship remains an empirical question. Accurate recall, contextual relevance, responsiveness to correction, and the user’s experience of being understood all deserve evaluation.
Memory also creates responsibilities. People change, and an account that once seemed appropriate may later feel incomplete or misleading. A reliable companion needs to leave room for users to correct records, reinterpret experiences, and decide what should continue to inform future dialogue. Remembering should support a person’s developing story without fixing them permanently to an earlier version of themselves.
From our paper to HuBan, the practical question is how AI can become a more dependable companion within appropriate limits: maintaining useful continuity, acknowledging uncertainty, supporting human relationships, and making room for professional help when needed.
I would welcome discussion of the following questions:
-
What should count as evidence that an AI meaningfully remembers a person?
-
Which aspects of autobiographical continuity matter most for companionship?
-
How should users retain authority over the memories and interpretations that shape later dialogue?
-
When does persistent interaction support companionship, and when might it encourage unhelpful dependence?
-
How can AI companions recognize the limits of their role and help people connect with human support?
These questions connect the organization of memory with the experience of companionship and the responsibilities that accompany it.