MENTI - סימולציה

The architecture

Three coordinated agents, in the same conversation.

Most AI "simulated patients" are a single role on the other end of a chat. MENTI runs three roles at once — a patient who resists when they should and opens when met with the right kind of attention, a supervisor who surfaces teaching moments in real time, and an evaluator who returns structured, dimension-by-dimension feedback when the session ends.

Agent 01

The Persona The patient, parent, or child

A clinically structured character built from a 10,000–30,000-word prompt with explicit psychological layers, a non-linear stage map, and a library of body-language cues.

Agent 02

Dr. Menton The real-time supervisor

A second voice that runs in parallel and surfaces a coaching note on every other turn — the "you might notice…" whisper that a good clinical instructor would give.

Agent 03

The Evaluator Post-session feedback

A structured analysis at the end of the session, scoring multiple clinically meaningful dimensions and returning the kind of written feedback a supervisor would actually use.

Three programmes

From clinical training to parents at home.

MENTI is one platform with three programmes, each led by a different research group.

MentiTherapist

Training therapists with realistic patients.

Clinical trainees practise with characters drawn from DSM categories — depression, BPD, complex grief, OCD, addiction — under a built-in clinical supervisor, with a structured post-session report. The current multi-site study follows clinical-psychology trainees through the Unified Protocol.

פורטרט מצויר של רמי — דמות מסימולציות MENTI

פורטרט מצויר של מאיה — דמות מסימולציות MENTI

MentiParent

Helping parents read their children.

Parents and family practitioners practise difficult conversations with simulated children, with a focus on parental reflective functioning — the ability to recognize and respond to the mental states behind a child's behaviour.

MentiCare

Holding steady when a child cannot.

Foster-care supervisors and professionals train on culturally specific simulations of foster-family dynamics, drawing on mentalization-based and trauma-informed frameworks.

פורטרט מצויר של אורית — דמות מסימולציות MENTI

Published evidence

MentiParent (Scientific Reports, in press): N = 96. Practitioners rated the simulator M = 5.61 / 7 for feasibility and M = 5.56 / 7 for acceptability; parents rated it M = 6.03 / 7 and M = 5.53 / 7. Cronbach's α > .94 on the MentiTherapist self-efficacy scales.

Outcome instruments: MentiParent uses 1–7 feasibility / acceptability items (PRFQ-18 planned for the multi-session follow-up). MentiTherapist uses MENTI-internal self-efficacy and feedback scales (α > .94); validated alliance and competence instruments planned. MentiCare uses field-tailored instruments; no canonical benchmark exists for foster-care supervision.

All published evidence to date measures trainee competence and feasibility. Downstream patient outcomes are the next research phase, not a current claim.

Voices from the study

What participants said.

At first I wondered how it would feel authentic when I was asked to play the role of someone else, but within seconds I channelled my own connection with my own child, what I feel with him, and was all in — emotionally.

— Practitioner participant, MentiParent feasibility study (N=96)

I liked getting the supervisor's suggestions, and they softened me towards Danny and his internal experience despite the external appearance.

— Practitioner participant, MentiParent feasibility study (N=96)

Very surprising. I have no children, and yet I cried several times during the exercise.

— Practitioner participant, MentiParent feasibility study (N=96)

How a session works

A complete training session in about ten turns.

Every MENTI session follows the same four-step rhythm. The training comes from the cycle, not from any single message.

Step 01

Choose a character.

Each persona is a hand-painted watercolour portrait paired with a structured clinical brief — age, diagnosis, presenting concern, training framework.

Step 02

Lead the conversation.

The persona responds in character — guarded, ambivalent, occasionally non-cooperative — exactly as a real patient might. The character won't surrender insight you haven't earned.

Step 03

Notice the supervisor.

Dr. Menton — a second AI voice — speaks every other turn with a coaching note. It is not the patient's voice, and the trainee can read or skip it freely.

Step 04

Receive structured feedback.

At the end, an evaluator agent returns a written analysis and a radar chart across multiple clinically meaningful dimensions — the kind of feedback an instructor would write.

Methodology

Clinical Prompt Engineering.

Most published AI-patient systems wrap a one-paragraph case description around a base model, and the resulting character agrees too readily, repeats themes, and breaks the moment it is pressed. We took the opposite bet. Through a method we call Clinical Prompt Engineering (CPE), we encode layered psychological architecture, contingency rules that link interactional events to internal state changes, mandated non-linear emotional trajectories, and boundary protocols into prompts that run 11,000 to 33,000 words.

Mentalization theory holds that epistemic trust — the willingness to treat another's communication as personally relevant and worth learning from — develops only inside relationships that reliably hold one's mind in mind. CPE asks whether a simulation built with that level of clinical depth can support the same kind of trust well enough to make rehearsal meaningful.

The result is a training conversation that holds its shape under pressure: the character resists when it should, opens when met with the right kind of attention, and the post-session report reads like one a clinical instructor would actually write. The same methodology is now being tested across four populations at two universities.

The falsifiable hypothesis behind CPE: that prompts of this length and structure produce character behaviour expert clinicians cannot distinguish from real clinical case formulations. Article 13 (Prompt Compression) tests it directly through a blinded expert evaluation of 33,000-word vs 3,000-word prompts across 14 clinically meaningful dimensions.

Built with

System card.

Model

Google Gemini (gemini-3-flash-preview, with gemini-3-pro-preview for the post-session analyst)

Hosting

Google Vertex AI · region me-west1 · under a Business Associate Agreement (effective 21 December 2025)

Protocol

Three independent prompt configurations invoked on the same growing transcript — sequenced, not concurrent. The Persona is called each turn; Dr. Menton on a configurable cadence; the Analyst once at session end. The agents share the transcript, not latent state.

Published research

A research pipeline, not a single paper.

One paper accepted at Scientific Reports, three more under peer review, and eleven in active preparation across the three programmes.

In press

Scientific Reports

Yirmiya, K.*, Refoua, E.*, Truscott, A., Reeve, H., Fonagy, P., & Elyoseph, Z. The MentiParent chatbot: An artificial intelligence-based approach to enhancing parental reflective functioning. (*equal contribution)

Under peer review

Child Abuse & Neglect

Elyoseph, T., Refoua, E., Novis-Deutsch, N., Yirmiya, K., Fonagy, P., & Enosh, G. Fostering reflection: development and initial evaluation of a mentalization-based GenAI simulator for foster-care supervisors.

Under peer review

CLPsych 2026 (workshop submission)

Holzman, Y., Rafaeli, E., Elyoseph, Z., Haber, Y., Yirmiya, K., Linkovski, O., Elyoseph, T., & Refoua, E. Clinical Prompt Engineering: encoding clinical knowledge into AI training simulations — a crisis-deployment case study.

In preparation

12 more articles, across the programmes

Highlights include: a within-subject test of AI-simulation training across three Unified Protocol modules (target: NPJ Digital Medicine); a blinded expert evaluation of 33,000-word vs 3,000-word clinical prompts; a randomised multi-session study of parental reflective functioning; and a two-study evaluation of Socratic questioning skill acquisition through AI-simulated patients (OSF pre-registered).

Team

A small team across three institutions.

MENTI is led by Elad Refoua at Bar-Ilan University, with co-investigators across Israeli universities and programme leads who run the three sub-projects.

Elad Refoua

Project Director

Directs the day-to-day research programme. Bar-Ilan University.

Prof. Eshkol Rafaeli

Co-PI

Bar-Ilan University · Psychotherapy and schema therapy.

Dr. Omer Linkovski

Co-PI

Bar-Ilan University · Psychopathology and digital methods.

Prof. Zohar Elyoseph

Co-PI

University of Haifa · AI in psychology.

Dr. Karen Yirmiya

Co-PI · MentiParent lead

UCL / Ben-Gurion University.

Yuval Holzman

Research coordinator

Bar-Ilan University · Coordinates the multi-site studies and clinical-validation pipeline.

Tal Elyoseph

MentiCare lead

Foster-care simulation research.

Senior theoretical collaboration

Prof. Peter Fonagy of UCL and the Anna Freud Centre — founder of Mentalization-Based Treatment — collaborates with the project on the theoretical framing of MentiParent. Prof. Fonagy is not a grant co-investigator and is not operationally involved in development.

Responsible AI

Built for clinical research, not for crisis support.

MENTI is a research and training environment. Below is how we handle session data, what the system does if a moment crosses a clinical safety line, and what the simulations are — and are not — built from.

Session transcripts

Session transcripts are stored in Google Cloud Storage in region me-west1, encrypted at rest with AES-256-GCM. They are retained for the duration of the IRB-approved study (default 365 days, configurable per study) and accessible only to authorised researchers behind multi-factor authentication. Transcripts are not used to retrain Gemini or any other model — Vertex AI's Business Associate Agreement prohibits this. Participants can request deletion at any time via a personal access code.

Crisis and escalation

Every simulation prompt includes a crisis protocol. If the trainee enters language indicating risk of harm to self or others — or if the simulated character is steered toward content that crosses clinical safety lines — the session is interrupted with a referral to human help (in Israel: ERAN 1201; NATAL *6768). MENTI is a training environment, not a crisis service.

Prompt provenance

The 11,000–33,000-word prompts behind each character are clinician-authored composites, not real patient material. They are built from clinical experience, published case formulations, and theoretical literature. No real-patient transcript, record, or identifying detail enters any MENTI prompt. Every prompt is reviewed by at least one expert clinician trained in the relevant framework before deployment.

Acceptable use

MENTI is built as a research and training environment for clinicians-in-training, parents, and caregivers practising skills they will later use with real people. It is not a therapy service, not a diagnostic tool, not a crisis-support line, and not a substitute for professional supervision or treatment. Every introductory screen names MENTI as an AI simulation.

Bias and fairness

A formal fairness audit of MENTI's simulations has not yet been conducted. The research roadmap includes a structured evaluation of how prompt content, character demographics, and trainee gender interact with model behaviour. In the interim, every simulation goes through expert clinical review with explicit attention to stereotype risk, and every prompt is gender-inclusive in Hebrew at the grammatical level.

Built carefully

Research-grade compliance, by design.

MENTI runs on Google Vertex AI under a Business Associate Agreement and is reviewed and approved by the Bar-Ilan University IRB. Data subject rights are exercised through a per-user access code.

GDPR-compliant

EU Data Protection Addendum in force; access, rectification, erasure, and portability supported via per-user access code.

Israeli Privacy Law

Compliant with the Privacy Protection Law (5741-1981) and the Privacy Protection Regulations (Data Security), 2017.

BIU IRB approved

Reviewed and approved by the Bar-Ilan University Institutional Review Board, Approval #290125444 (29 January 2025).

Funded by Israeli MOST

Israeli Ministry of Innovation, Science & Technology — Proposal Ref. 0008682.

Encrypted at rest

AES-256-GCM encryption for session data; HTTPS/TLS for transport; structured audit logs in Google Cloud Logging.

In preparation

Try a sample simulation.

A short public training demo in English and Hebrew is in preparation. Leave your email below and we will notify you when it is available.

Register interest

Contact

For institutions, researchers and press.

If you are evaluating MENTI for a training programme, a research collaboration, or a press feature, this form goes straight to Elad.

Already have an access code?

If you finished a previous MENTI session and saved your personal access code, you can view, export, or delete your data.

Three voices in every conversation.

Three coordinated agents, in the same conversation.

The Persona The patient, parent, or child

Dr. Menton The real-time supervisor

The Evaluator Post-session feedback

From clinical training to parents at home.

Training therapists with realistic patients.

Helping parents read their children.

Holding steady when a child cannot.

What participants said.

A complete training session in about ten turns.

Choose a character.

Lead the conversation.

Notice the supervisor.

Receive structured feedback.

Clinical Prompt Engineering.

System card.

A research pipeline, not a single paper.

A small team across three institutions.

Built for clinical research, not for crisis support.

Session transcripts

Crisis and escalation

Prompt provenance

Acceptable use

Bias and fairness

Research-grade compliance, by design.

GDPR-compliant

Israeli Privacy Law

BIU IRB approved

Funded by Israeli MOST

Encrypted at rest

Try a sample simulation.

For institutions, researchers and press.

Already have an access code?

ברוכים הבאים לMENTI

פידבק מקצועי

📋 הנתונים שלי

Three coordinated agents, in the same conversation.

The Persona The patient, parent, or child

Dr. Menton The real-time supervisor

The Evaluator Post-session feedback

From clinical training to parents at home.

Training therapists with realistic patients.

Helping parents read their children.

Holding steady when a child cannot.

What participants said.

A complete training session in about ten turns.

Choose a character.

Lead the conversation.

Notice the supervisor.

Receive structured feedback.

Clinical Prompt Engineering.

System card.

A research pipeline, not a single paper.

A small team across three institutions.

Built for clinical research, not for crisis support.

Session transcripts

Crisis and escalation

Prompt provenance

Acceptable use

Bias and fairness

Research-grade compliance, by design.

GDPR-compliant

Israeli Privacy Law

BIU IRB approved

Funded by Israeli MOST

Encrypted at rest

Try a sample simulation.

For institutions, researchers and press.

Already have an access code?

ברוכים הבאים לMENTI

פידבק מקצועי