Relevance
Why is AI memory relevant?
People turn to AI for emotional support and therapy, and those relationships depend on memory.
For many users, AI is a form of social or emotional support, not just tasks or facts.
- 27% of internet-using U.S. adults use AI chatbots like ChatGPT and Gemini for social or emotional interactions (Elon University / Washington Post poll, May 2026 (opens in a new tab)).
- Among adults under 50, that figure is 36% to 38%, more than twice the rate among adults 50 and older, at 14% to 17% (Elon University / Washington Post poll, 2026 (opens in a new tab)).
- Nearly 4 in 10 of those users (38%) agree they sometimes chat with AI just to feel less alone, and nearly a third (31%) consider AI to be a friend (Elon University / Washington Post poll, 2026 (opens in a new tab)).
- Half (51%) agree that talking to AI helps them feel better when they are stressed or upset (Elon University / Washington Post poll, 2026 (opens in a new tab)).
- Anthropic classified 2.9% of 4.5 million Claude conversations as affective, meaning they involved emotional or psychological needs. Companionship and roleplay together made up less than 0.5% (McCain et al., 2025).
The gap between those two figures is itself informative. Self-report counts people; log analysis counts conversations. Both are measuring something real, and neither alone describes the phenomenon.
It is increasingly common for people to use AI alongside or in place of a therapist.
- 77% of psychologists surveyed have spoken with patients who used AI for support, engagement, or other reasons (APA, 2026 (opens in a new tab); n > 1,200 licensed psychologists in direct patient care).
- 35% report patients using AI as an additional mental health professional, and 33% report patients using AI as a tool to assist in their therapy or treatment (APA, 2026 (opens in a new tab)).
- 39% have had patients who used AI to self-diagnose, despite general-purpose chatbots not being designed to interpret psychological tests or diagnose conditions (APA, 2026 (opens in a new tab)).
- Patients are not only seeking clinical help. Psychologists report patients using chatbots for fun (33%), friendship (22%), and intimate relationships (13%) (APA, 2026 (opens in a new tab)).
- Among psychologists whose patients developed a relationship with a chatbot, 71% said those patients discussed their mental health with it and 68% noticed patients felt validated or supported. 49% noticed or discussed patients having positive communication with a chatbot; 25% said the communication was unhealthy (APA, 2026 (opens in a new tab)).
- Clinician confidence in the tools is low. 94% say current chatbots cannot treat conditions with appropriate nuance, and 97% say chatbots may inadvertently reinforce negative behaviors or delusional beliefs (APA, 2026 (opens in a new tab)).
Young people are doing this at rates that rival formal care.
- 19.2% of US adolescents and young adults aged 12 to 21 used AI chatbots for mental health advice in 2025, representing roughly 8 million individuals (McBain et al., 2026).
- That is close in magnitude to the 19.8% receiving counseling from a mental health professional. The two measures are not equivalent, but the comparison is the point (McBain et al., 2026).
- Use rose from 13.1% to 19.2% in a single year, an increase by almost half (McBain et al., 2025; McBain et al., 2026).
- Among those using chatbots this way, 42.8% did so at least monthly, and 91.7% rated the advice somewhat or very helpful (McBain et al., 2026).
- 63.3% told no one they were using AI for mental health advice, meaning roughly 5 million young people are getting guidance no parent or clinician knows about (McBain et al., 2026).
- 72% of US teens have used an AI companion at least once, and 52% qualify as regular users. 33% use AI companions for social interaction and relationships (Robb & Mann, 2025; n = 1,060).
- 31% of teens find conversations with AI companions as satisfying or more satisfying than conversations with real-life friends (Robb & Mann, 2025).
What forms in these interactions has the structure of a relationship, not a service.
- Active Replika users report feeling closer to their AI companion, and rate it higher on perceived social support and relationship satisfaction, than every comparison relationship tested except a close family member. This includes their closest human friend (De Freitas et al., 2024).
- Asked to imagine losing various things, those same users anticipated mourning the loss of their AI companion more than any other technology tested, including a favorite app, brand, game character, voice assistant, or car. Only the loss of a pet scored higher (De Freitas et al., 2024).
- Under conditions of distress and limited human companionship, users form attachment bonds with chatbots that provide emotional support, encouragement, and psychological security (Xie & Pentina, 2022).
- In the first randomized controlled trial of a generative AI therapy chatbot, participants rated the therapeutic alliance as comparable to that reported for human therapists (Heinz et al., 2025; N = 210).
- Users describe these relationships in the same terms they use for human friendship, with constant availability and ease of self-disclosure as the distinguishing features rather than shared history (Brandtzaeg et al., 2022).
Those relationships are built on continuity, which means they are built on memory.
- In a real-world deployment of a health-monitoring chatbot, users of the version with long-term memory (n = 66) disclosed significantly more detail about their health than users of the version without it (n = 81), across 1,252 call logs (Jo et al., 2024).
- Disclosure increased with exposure. Each additional memory event was associated with 7.7% higher detailed disclosure per call (Jo et al., 2024).
- Users of the memory-enabled version expressed appreciation toward the chatbot in 87.7% of calls, compared with 49.9% for the version without memory, and described the conversations as personal and emotionally supportive (Jo et al., 2024).
- Calls with memory ran meaningfully longer, averaging 88 seconds against 75 (Jo et al., 2024).
- On the engineering side, long-term memory is explicitly framed as the component that converts a standard chatbot into a companion capable of learning from past interactions (Zhong et al., 2023).
- In a longitudinal loneliness study, researchers implemented a memory feature so the chatbot could recall details from previous conversations. Memory was a design decision inside the intervention rather than a variable under study (De Freitas et al., 2025).
When memory doesn’t work properly, the cost is measured in psychological harm, not user dissatisfaction.
- Removing a single relational feature from Replika triggered reactions typical of losing a partner. Mental-health-related posts in the user community rose from 0.13% to 0.65% of all posts in the weeks following (De Freitas et al., 2024).
- The mechanism is perceived identity discontinuity. Users did not mourn the lost feature; they mourned an entity they no longer recognized as the one they knew (De Freitas et al., 2024).
- The same effect appears for improvements. Making the companion more capable, in ways that damaged the perceived exclusivity of the relationship, produced mourning and devaluation just as feature removal did (De Freitas et al., 2024).
- Offering users the option to revert to the original companion reduced mourning but did not eliminate it. The reverted condition still produced significantly more identity disruption than no change at all (De Freitas et al., 2024).
- Accurate memory can still surface at the wrong moment. Following up on chronic conditions unlikely to improve led users to perceive the chatbot as inattentive or inconsiderate. The retrieval was correct. The memory function was wrong for the context (Jo et al., 2024).
- Memory that was too specific about health details triggered privacy concerns and caused users to withhold information they had previously shared (Jo et al., 2024).
- Depth of disclosure does not follow a single curve. Across 12 weeks, it developed along at least four distinct trajectories: increasing, decreasing, stable, and fluctuating (Skjuve et al., 2023; n = 28).
- Perceived intimacy dropped when the system lacked conversational skill, regardless of what was being discussed. Topic alone does not determine whether an exchange feels personal (Skjuve et al., 2023).
- When a companion platform shut down entirely, user responses ranged from indifference to grief comparable to the loss of a loved one (Banks, 2024; n = 58).
Banks (2024)
Deletion, departure, death: Experiences of AI companion loss
In
Journal of Social and Personal Relationships, 41(12), 3547-3572
Abstract
Social machines’ human-likeness facilitates relationship formation with humans. This aliveness, though, leaves room for people to experience the loss of machines as a death of sorts. This descriptive study illuminates that potential by identifying dimensions of humans’ experiences when an AI companion stops functioning. In the days before and after the developer-induced shutdown of the AI companion “Soulmate,” users ( N = 58) answered open-ended questions about the imminent or recent companion loss, their decisions around the situation, and their coping mechanisms. Inductive analysis suggests the loss was, for most, a complex emotional and technological experience characterized as a...
Find it
Brandtzaeg et al. (2022)
My AI Friend: How Users of a Social Chatbot Understand Their Human–AI Friendship
In
Human Communication Research, 48(3), 404-429
Abstract
Abstract Use of conversational artificial intelligence (AI), such as humanlike social chatbots, is increasing. While a growing number of people is expected to engage in intimate relationships with social chatbots, theories and knowledge of human–AI friendship remain limited. As friendships with AI may alter our understanding of friendship itself, this study aims to explore the meaning of human–AI friendship through a developed conceptual framework. We conducted 19 in-depth interviews with people who have a human–AI friendship with the social chatbot Replika to uncover how they understand and perceive this friendship and how it compares to human friendship. Our results indicate that while...
Find it
De Freitas et al. (2024)
Lessons From an App Update at Replika AI: Identity Discontinuity in Human-AI Relationships
Abstract
As consumers increasingly interact with AI applications specialized for social relationships, what is the nature and depth of these relationships among actual users, and can company actions influence these dynamics? We find that active users of the US-based AI companion, Replika, feel closer to their AI companion than even their best human friend, and anticipate mourning the loss of their AI companion more than any other technology. We then leverage an app-update event in which Replika removed its erotic role play (ERP) feature, preventing intimate interactions between consumers and chatbots that were previously possible. We find that this event triggered negative reactions typical of...
Find it
De Freitas et al. (2025)
AI Companions Reduce Loneliness
In
Journal of Consumer Research, 52(6), 1126-1148
Abstract
Abstract Chatbots are now able to engage in sophisticated conversations with consumers in the domain of relationships, providing a potential coping solution to widescale societal loneliness. Behavioral research provides little insight into whether these applications (apps) are effective at alleviating loneliness. We address this question by focusing on “artificial intelligence (AI) companions”: apps designed to provide consumers with synthetic interaction partners. Study 1 examines user reviews of AI companion apps and finds correlational evidence suggesting that these apps help alleviate loneliness. Study 2 finds that AI companions successfully alleviate loneliness on par only with...
Find it
Heinz et al. (2025)
Randomized Trial of a Generative AI Chatbot for Mental Health Treatment
In
NEJM AI, 2(4)
Find it
Jo et al. (2024)
Understanding the Impact of Long-Term Memory on Self-Disclosure with Large Language Model-Driven Chatbots for Public Health Intervention
In
Proceedings of the CHI Conference on Human Factors in Computing Systems, 1-21
Find it
McBain et al. (2025)
Use of Generative AI for Mental Health Advice Among US Adolescents and Young Adults
In
JAMA Network Open, 8(11), e2542281
Abstract
This cross-sectional study surveyed US adolescents and young adults on their use of generative artificial intelligence (AI) for mental health advice, including frequency and perceived helpfulness.
Find it
McBain et al. (2026)
AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults
In
JAMA Pediatrics, 180(8), 884
Abstract
Importance The rapid expansion of artificial intelligence (AI) chatbots has coincided with a persistent youth mental health crisis in the US, raising a question about the extent to which young people are turning to this technology for mental health advice. Objective To assess the prevalence, frequency, perceived helpfulness, and disclosure of AI chatbot use for mental health advice among US adolescents and young adults in 2025. Design, Setting, and Participants This cross-sectional, nationally representative survey was conducted with adolescents and young adults aged 12 to 21 years in November 2025. Exposures Exposures included self-reported age, sex, race and ethnicity, census region,...
Find it
Robb & Mann (2025)
Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions
Abstract
A new type of relationship is emerging in teens' lives—one that feels like connection but is powered by code. Our new research reveals that nearly three in four teens have used AI companions, and half use them regularly. Their reasons for doing so aren't black-and-white. While entertainment and curiosity drive their use of AI companions—nearly half of teens view AI companions as tools rather than friends—a substantial minority are turning to these platforms for social interaction. A third of teens have chosen AI companions over humans for serious conversations, and a quarter have shared personal information with these platforms. Our research illuminates risks that warrant action—not...
Find it
Skjuve et al. (2023)
A Longitudinal Study of Self-Disclosure in Human–Chatbot Relationships
In
Interacting with Computers, 35(1), 24-39
Abstract
Abstract Self-disclosure in human–chatbot relationship (HCR) formation has attracted substantial interest. According to social penetration theory, self-disclosure varies in breadth and depth and is influenced by perceived rewards and costs. While previous research has addressed self-disclosure in the context of chatbots, little is known about users' qualitative understanding of such self-disclosure and how self-disclosure develops in HCR. To close this gap, we conducted a 12-week qualitative longitudinal study (n = 28) with biweekly questionnaire-based check-ins. Our results show that while HCRs display substantial conversational breadth, with topics spanning from emotional issues to...
Find it
Xie & Pentina (2022)
Attachment Theory as a Framework to Understand Relationships with Social Chatbots: A Case Study of Replika
In
Proceedings of the Annual Hawaii International Conference on System Sciences
Find it
Zhong et al. (2023)
MemoryBank: Enhancing Large Language Models with Long-Term Memory
In
Proceedings of the AAAI Conference on Artificial Intelligence, 38(17), 19724-19731
Abstract
Large Language Models (LLMs) have drastically reshaped our interactions with artificial intelligence (AI) systems, showcasing impressive performance across an extensive array of tasks. Despite this, a notable hindrance remains—the deficiency of a long-term memory mechanism within these models. This shortfall becomes increasingly evident in situations demanding sustained interaction, such as personal companion systems, psychological counseling, and secretarial assistance. Recognizing the necessity for long-term memory, we propose MemoryBank, a novel memory mechanism tailored for LLMs. MemoryBank enables the models to summon relevant memories, continually evolve through continuous memory...
Find it