MindfulAgents: Personalizing Mindfulness Meditation via an Expert-Aligned Multi-Agent System

Paper Title

MindfulAgents: Personalizing Mindfulness Meditation via an Expert-Aligned Multi-Agent System

Publication Info

• Topic area: Personalization and engagement in digital mindfulness meditation using AI-driven systems.
• Keywords: Mindfulness meditation, personalization, multi-agent system, large language models, user engagement, mental health, reflection, expert alignment, digital interventions, meditation apps.

Background and Problem

• Problem / challenge: Sustaining user engagement in mindfulness meditation apps is a persistent challenge, with high attrition rates (95% within the first month for many apps). Existing solutions lack scalable personalization and often fail to address long-term engagement.
• Significance: Mindfulness meditation is an evidence-based practice for improving mental health, but its benefits are contingent on sustained engagement. Addressing this gap can make meditation more accessible and impactful for diverse populations.
• Motivation and related work: Prior work has demonstrated the potential of large language models (LLMs) for personalization in mental health contexts but has faced challenges with reliability, safety, and long-term engagement. Current systems often focus on single-session experiences or static content, leaving a gap in scalable, personalized, and expert-aligned solutions.

Solution

• Proposed approach: MindfulAgents, a multi-agent system powered by LLMs, delivers expert-aligned, personalized mindfulness meditation guidance by integrating three components: an Expert-Alignment Agent, a Reflection Agent, and a Personalization Agent.
• Novelty:
  1. Combines expert-aligned safety templates with real-time personalization for scalable, reliable meditation guidance.
  2. Introduces a Reflection Agent to foster self-awareness and deepen user engagement through interactive dialogue.
  3. Demonstrates long-term engagement and mindfulness improvements through a multi-agent architecture evaluated in both lab and field studies.
  4. Provides a modular framework adaptable to various mindfulness curricula.
• Procedure and key techniques:
  1. Expert-Alignment Agent: Generates safety meditation templates grounded in the Unified Mindfulness (UM) framework using supervised fine-tuning (SFT) and expert feedback.
  2. Reflection Agent: Engages users in pre-session reflection on their current state, past sessions, and mindfulness concepts, leveraging retrieval-augmented generation (RAG).
  3. Personalization Agent: Adapts expert-aligned templates in real-time based on user inputs (mood, goals, history, etc.) and reflective dialogue.
  4. System evaluated through iterative co-design with mindfulness experts and two user studies (lab and deployment).

Results

• Concrete findings:
  • In a lab study (N=13), MindfulAgents significantly improved in-session engagement (p = 0.011), self-awareness (p = 0.014), and stress reduction (p = 0.020).
  • In a 4-week deployment study (N=62), MindfulAgents achieved higher long-term engagement (0.94 vs. 0.40 sessions/day, p = 0.002) and greater mindfulness improvements (p = 0.023) compared to a static baseline.
• Advantage over baselines:
  • Outperformed static and partially personalized baselines in engagement, user preference, and mindfulness outcomes.
  • Users preferred MindfulAgents for its relevance, diversity, and interactive reflection features.
• Experiments / evaluation:
  • Lab study: Within-subject design comparing three conditions (StaticAgent, PersonalAgents, MindfulAgents) on engagement, personalization, and stress reduction.
  • Deployment study: Between-subject design comparing StaticAgent and MindfulAgents over 4 weeks, measuring engagement, mindfulness (FFMQ-SF), mood (PANAS-SF), anxiety (GAD-7), and sleep quality (PSQI).
• Limitations and future work:
  • Short deployment duration (4 weeks) limits assessment of long-term psychological benefits.
  • Latency in LLM processing and lack of personalized audio delivery hindered user experience.
  • Future work should explore longitudinal studies, passive sensing for reduced user input effort, and dynamic evolution of the system’s role over time.

Summary

MindfulAgents is a multi-agent system that combines expert-aligned safety, interactive reflection, and real-time personalization to enhance engagement in mindfulness meditation. Evaluations showed significant improvements in user engagement, mindfulness, and stress reduction compared to static baselines. The system’s modular design supports adaptability to various mindfulness frameworks, offering a scalable solution for personalized digital meditation. Future research should address technical limitations and explore long-term impacts on mental health.