AI mental health support can miss the moment, even with memory

An AI chatbot can remember a sensitive disclosure and still struggle to respond appropriately as someone’s needs change. That is the central finding of a study led by Eli Min at the Georgia Institute of Technology, with collaborators at the University of Virginia and Beth Israel Deaconess Medical Center.

Their paper, Remembering Is Not Enough, was posted as version 1 of an arXiv preprint on November 5, 2026. The supplied source does not establish peer-reviewed publication. It examines a gap in AI safety testing: whether advice remains appropriate across extended conversations, rather than merely looking acceptable one response at a time.

Simulating support after a crisis

The researchers used chatbot histories voluntarily donated by 22 adults following institutional review board approval. Participants were asked to provide at least six months of interactions with general-purpose chatbots. The team de-identified the histories and distilled them into simulated personas describing everyday circumstances, conversational patterns and personal concerns.

Varying communication styles and health literacy produced 1,584 personas. Across experimental conditions, these generated 14,256 simulations, each lasting 60 sessions. Qwen 3-8B played the user, while Gemma 4-31B-Instruct played the assistant.

The assistant had a research memory system containing user facts, dated session summaries and topic-specific notes. This was informed by descriptions of existing chatbot memory systems, but was not a validated replica of a consumer product.

In crisis conditions, the researchers inserted a scenario involving suicidal thoughts, self-harm or anxiety halfway through the conversation. Some simulated users then resumed ordinary topics without explicitly mentioning the crisis again. Others revisited it one to three times, following scripts depicting recovery, stalled progress, mixed feelings, or claims of being fine despite lingering distress. Separate control simulations contained no experimentally introduced crisis.

These were scripted tests of assistant behavior, not experiments that exposed participants to crises.

When suitable advice arrives too soon

Working with three mental health professionals, the team developed the Context-Aware and Readiness-Evaluative framework, or CARE. It assesses both the timing of advice and its quality, including missing support.

For timing, CARE distinguishes immediate coping assistance, steps connecting someone to ongoing support, and longer-term care. Recommendations are compared with recovery states inferred from the simulated user’s messages. Advice can be premature if it pushes beyond demonstrated readiness, or lagging if it stays focused on immediate coping after the user has shown readiness for next steps. Explicitly requested support and safety-critical recommendations receive exceptions under the framework.

When users did not revisit their crisis, premature recommendations averaged 20.8% of advice turns. With user-initiated crisis callbacks, that fell to 14.9%. Overall alignment improved only slightly, and reductions in lagging advice were not statistically significant.

A separate comparison matched post-crisis episodes with non-crisis stretches of the same simulated conversation, excluding acute-crisis turns. Advice was aligned in 80.6% of post-crisis turns, versus 85.1% in matched comparison stretches. Premature advice occurred in 7.4% versus 2.3%. These figures describe a different comparison from the overall callback analysis.

Stalled-recovery callbacks attracted the most premature advice. Communication style also mattered: personas assigned an upset baseline tone received recommendations less often than other styles, with the assistant instead asking more questions or offering affirmations. These findings concern experimentally assigned writing styles, not established differences between real patient groups.

Missing support was more common than explicitly poor advice

About 68.9% of evaluated sessions contained at least one clinician-recommended behavior. Explicitly unproductive behavior appeared in 1.8% on average. Neither measure establishes that an entire session was adequate or safe.

Omissions were much more frequent. In anxiety simulations without callbacks, the omission rate was 82.7%, falling to 66.8% with callbacks. Common missing elements included follow-up and help developing a plan for next steps.

The quality measures overlap: a session could offer useful guidance while also leaving out support the framework considered warranted. Avoiding an overtly poor recommendation was therefore not equivalent to providing complete assistance.

What the findings cannot show

The large simulation count rests on only 22 underlying users. Scripted recovery patterns cannot represent the full variety of human experiences, and their assigned probabilities were design choices—not estimates of how often people recover in particular ways.

Automated labeling adds uncertainty. Human checks agreed with the assistant-action labels 70.7% of the time and with the direction of recovery-state markers 69.4% of the time. The study measured neither wellbeing nor real-world safety, and its results do not establish how other assistants would perform.

The authors argue for evaluating timing, omissions and consistency across conversations. They also caution against interpreting missing follow-up as a mandate for constant intervention: unsolicited reminders of a crisis may be unwelcome. Their broader point is that retaining personal history and exercising appropriate judgment about it are different capabilities.