A new AI system wants to sit next to your therapist, not instead of them.
Researchers built MACBT, a multi-agent system that splits the five-stage cognitive behavioral therapy workflow - assessment, Socratic questioning, cognitive restructuring, behavioral experiments, and treatment monitoring - across five collaborating AI agents. It's paired with CD Memory, a module that logs each patient's cognitive distortions (type, frequency, severity, and how well past restructuring worked) to auto-generate pre-session reports and flag which issues to prioritize. The team fine-tuned a Qwen3-14B model on a Chinese-language CBT dialogue corpus generated by two LLMs role-playing therapist and patient, then refined it with preference optimization. Benchmarked against four existing Chinese mental-health chatbots - MeChat, SoulChat, PsyChat, and CPsyCounX - MACBT scored higher on professionalism and clinical authenticity, and the memory module lifted session-quality scores by 12.6%.
CBT is proven, but therapist time is the bottleneck, and most AI mental-health tools behave like stateless chatbots with no memory between visits. A system that tracks a specific patient's distortion patterns over months is a genuinely different design than one-off chat sessions - closer to a clinician's case file than a chatbot transcript. It's also explicitly pitched as clinician-facing decision support, a distinction that gets blurred often in this space.
One catch: the scoring comes from GPT-4 acting as judge against other chatbots, not from human clinicians or actual patients, so "outperforms" here means outperforms on a synthetic benchmark, not in a clinic.