Persistent memory is becoming a default capability in consumer AI systems. OpenAI, Google, and Anthropic have each introduced mechanisms that let their models carry information across conversations, remembering preferences, past disclosures, and stated goals without being told again. In most consumer contexts, this is a convenience. In mental health contexts, it is a design decision with clinical implications the field has barely begun to examine.

The question is not whether memory improves continuity. It almost certainly does. The question is whether the systems doing the remembering are equipped to use that memory safely.

Why continuity matters

In psychotherapy, memory is not incidental to the work, it is foundational to it. A clinician recalls what a client said three weeks ago, notices when their account of an event has shifted, tracks the relationship between current distress and earlier history. Research on the therapeutic alliance in digital tools finds that continuity, including context-aware responses that reflect prior disclosures, is among the key mechanisms that build relational bonds between users and AI systems (Xu et al., 2025).

Digital mental health tools have historically lacked this. Each session starts from scratch, and users repeat themselves. Persistent memory addresses a real structural weakness, and that is worth acknowledging before the risks are examined.

Where AI memory differs from clinical memory

Large language models generate responses by predicting the next token given prior context. When a model retrieves a stored memory, it is not recalling an event the way a therapist does. It is using a text string as part of a prediction, and that string may be accurate, outdated, misrepresented in storage, or subtly distorted in a way neither the user nor the system can flag. A stored note that a user "tends toward catastrophising", surfaced automatically in every later exchange, is doing something that would ordinarily require clinical judgement, and it is doing it probabilistically.

What the evidence shows

The picture is encouraging in general terms and thinner where it matters most. A meta-analysis of 176 randomised controlled trials found mental health apps produce meaningful, small-to-moderate reductions in depression and anxiety symptoms compared with inactive controls (Linardon et al., 2024). A systematic review of large language models in mental health found promising performance in detection and destigmatised support, but concluded that the risks of clinical deployment may currently outweigh the benefits, citing the absence of a benchmarked ethical framework (Guo et al., 2024).

Research on memory specifically is thinner still. Anthropomorphic design features, including conversational continuity, shape user trust and dependency in ways that can run in either direction (Monteith et al., 2025). Overreliance on AI recommendations is already a documented risk in mental health decision-making, and a system that references stored memory as though it were verified fact, rather than a stored approximation, invites a level of trust it has not earned (Auf et al., 2025).

The specific risks

Four follow directly from what memory enables. Stored disclosures do not update themselves, so a user who has recovered from a crisis may still be treated by the system as though they have not. Models are unlikely to challenge a self-critical description once it is stored as fact, entrenching rather than examining it. And mental health disclosures are among the most sensitive categories of personal data: the EU AI Act introduces governance obligations for high-risk healthcare systems, but the gap between regulatory compliance and genuine user control over stored data remains wide (European Parliament and Council of the European Union, 2024).

What this means for builders

None of this argues against persistent memory. It argues for building it with constraints most current implementations lack. Users need to see, correct, and delete what is stored, not as a feature request but as a precondition for informed consent. Memory needs an expiry: a note about someone's state in one period of their life is not necessarily true months later. And anywhere memory intersects with risk, self-harm history, suicidal ideation, acute crisis, the product needs defined escalation protocols, not a retrieval index standing in for clinical judgement.

The harder, and more honest, design brief is a memory system that knows what it is: a text retrieval mechanism inside a probabilistic model, not a surrogate clinician with a case history. That is a harder thing to ship. It is the right one to build.

🪺