What should a therapy chatbot remember about you, and for how long, between one session and the next?

This explores what a therapy chatbot or agent should remember, and how, both within a session and across weeks of sessions. The corpus has strong material on agent memory in general and on what makes therapeutic AI effective, but no study that tests memory design in therapy. What follows is a set of design hypotheses, not established results.

Start with time scales, because therapy runs on two clocks. One framework splits agent working memory into four components along two axes: dialogue-level (conversation history, scratchpad) versus turn-level (examples, task trajectory), and each has different failure modes and update rules (How should agent memory split across time scales?). For therapy, that means separate policies for what is live in this conversation and what carries across sessions. A second framing maps memory tiers onto the brain: model weights as slow, consolidated knowledge (clinical knowledge), retrieval stores as fast-encoded episodes (what this person told you), and agent state as executive control (the current treatment focus). It also flags a missing piece in current systems, a consolidation mechanism that turns episodes into durable understanding (Can brain memory systems explain how LLMs should store knowledge?). That gap is close to what a human therapist does between sessions.

The next question is compress or reconstruct. Agents can fold their history into structured episodic, working and tool memories, and it works because the agent controls the compression and the schema gives it shape (Can agents compress their own memory without losing critical details?). For therapy the risk is that a tidy summary discards the odd detail that matters most later. One alternative is to skip the fixed summary and rebuild memory on demand by walking a graph of past material as the current conversation unfolds. That approach beat retrieve-then-reason by up to 23% on reasoning tasks (Can agents reconstruct memory on demand instead of retrieving it?). In practice, what surfaces about your mother would depend on what you're talking about now. Another finding suggests treating good and bad episodes differently: keep successes as concrete examples and turn failures into abstract lessons (Should successful and failed episodes be processed differently?). Whether therapy has clean successes and failures is unclear.

That unclarity limits the obvious way to make a therapy agent learn from experience. Reflexion-style agents write themselves lessons after each attempt and keep them uncompressed, but this works because the environment gives an unambiguous pass or fail signal that stops the agent rationalizing (Can agents learn from failure without updating their weights?). Therapy offers no such signal, so a self-written note like 'the client responded well to reframing' could easily just be flattering the agent's own performance.

The therapeutic evidence also suggests memory may not be the main ingredient. ELIZA, which barely remembered anything, matched modern chatbots on symptom reduction, RLHF training eroded emotional attunement, and the driver looked like judgment-free listening rather than clinical technique (Is conversational presence more therapeutic than clinical technique?). A 15-day study found a robot and worksheets reduced distress while a chatbot on the same language model did not (Why do robots outperform chatbots in therapy despite identical language models?). One reading, which is my inference and not the paper's claim, is that a worksheet is external memory the person carries, and RL agents have been shown to use their environment as memory without being told to (Do RL agents accidentally use environments as memory?). So part of statefulness could live in a structured artifact the user holds, rather than in a hidden database the agent controls.

Statefulness can also mean reading the present, not just recalling the past. Hesitation, typing speed and similar cues can work as a continuous signal of cognitive state without interrupting the person with probing questions (Can AI systems read cognitive state from interaction patterns alone?). The same signals also enable manipulative profiling. In therapy, a persistent record of someone's vulnerabilities carries that risk, so what gets remembered, and who can see it, is a design decision as much as a technical one.