Toward a Resonant AI: Modeling the Dynamics of Mind through Buddhist Texts

Abstract

This work began with a question: can an AI system be shaped to resonate with the subtle transformations of consciousness described in Buddhist contemplative traditions? Could it be used not simply to classify or summarize, but to map, interpret, and reflect the transitions in mental states—such as those described in the Bardo Thödol, which depicts consciousness in intermediate states between life, death, and rebirth?

Our answer is the development of a proof-of-principle system: a meaning-driven, Buddhist-informed RAG pipeline designed to describe and analyze transitions in mind. This system leverages vector embeddings of Dharma texts and uses large language models to analyze the symbolic, emotional, and energetic dimensions of the text, producing structured descriptions of mental dynamics. The result is a black-box cognitive model capable of modeling meaningful transformations within inner experience.

While deeply grounded in a Buddhist view of mind, the implications are broader: any domain involving transitions of mental states—from meditation and psychotherapy to political behavior and social emotion—might benefit from models that can track, interpret, and reflect mental state dynamics.

Methodology

Sources

The system currently incorporates a curated corpus of Buddhist texts:

The Bardo Thödol (Tibetan Book of the Dead)
Selected retreat and study materials on Mahamudra and symbolic Vajrayāna practices

Texts including the Ocean of Definitive Meaning, Wheel of Life, Spiral Path, Heart Sutra, Sutra of Golden Light, Abhidharma, and Dhammapada are queued for inclusion as the next phase of the enquiry. These will be processed using the same pipeline and worldview, enabling coherent integration with the current dataset.

Segmentation and Embedding

Text segments are processed through:

Segmentation into thematic or dynamic units.
Semantic embedding using SentenceTransformers, capturing the subtle relationships and symbolic patterns.
Storage in a Chroma vector database for high-fidelity retrieval.

Each segment is later processed to generate a structured JSON description of its mental dynamics.

Mental Dynamics Descriptor Format

Each text segment is analyzed for:

Cognitive and emotional tone
Symbolic field
Energetic mode
Archetypal role
Transition directionality
Subtle resonance

These are encapsulated as Mental Dynamics Descriptors (MDDs)—vector-compatible representations of mind states and their dynamics.

Query Process and Prompting Strategy

Queries are submitted to the system as open-ended questions designed to elicit resonance rather than a fixed answer. For example:

“Say something about advice given about the experience of fear in the Bardo and advice for the experience of everyday fear.”

These queries are passed through a RAG pipeline:

The query is embedded and compared to the vector store.
Relevant chunks are retrieved.
The full context is passed to an LLM (Mistral, GPT-3.5, or GPT-4o).
The LLM generates a synthesized response, guided by symbolic and cognitive context.

Prompts for generation and augmentation follow a carefully designed schema (see Appendix: Prompt Engineering).

Results: Comparative Model Analysis

Mistral (local model)

Provided structurally coherent answers.
Captured symbolic themes like death, yidam, projection.
Less sensitive to nuance; limited reflection.

GPT-3.5

Articulated psychological correspondences with more emotional framing.
Bridged the symbolic Bardo experiexnce and everyday emotional processes.

GPT-4o

Offered the richest response.
Integrated Buddhist philosophy, symbolic vision, and therapeutic insight.
Reflected archetypal structure with notable clarity.

Summary Table

Model	Archetypal Resonance	Psychological Framing	Conciseness	Nuance	Deployment
Mistral	✅ Basic	✅ Light	✅ High	⚠️ Moderate	✅ Local
GPT-3.5	✅ Moderate	✅ Solid	✅ Moderate	✅ Good	✅ API
GPT-4o	✅ Deep	✅ Rich	⚠️ Lower	✅ Excellent	⚠️ API

Key: ✅ Strong | ⚠️ Acceptable but limited

Prompt Engineering

Open-ended approach

Our queries emphasized resonance over instruction, inviting the model to reflect rather than merely explain. This openness proved critical in eliciting symbolic associations and layered interpretations.

Prompts used in generation and augmentation

Generator prompt (Mistral):

pythonCopyEditYou are a reflective cognitive agent analyzing the spiritual, emotional, and symbolic dimensions of a text segment from the Tibetan Book of the Dead.
...
Return a single valid JSON with cognitive_tone, emotional_tone, symbolic_elements, energetic_mode, agency, tension, transition_vector, archetypal_field, subtle_resonance, comment, and confidence.

Augmentor prompt (GPT-3.5):

pythonCopyEditYou are a second-pass analyst reviewing a Mistral-generated JSON descriptor of a Buddhist text segment...
Return:
{
  "gpt35_review": {
    "commentary": "...",
    "agreement_with_mistral": true or false,
    "additions": {...},
    "gpt35": true
  }
}

These prompts enabled both local and cloud-based models to contribute to an evolving resonant framework of mind interpretation.

Summary and Positioning

This project demonstrates the feasibility of creating a resonant AI using a RAG system grounded in a Buddhist worldview. The focus is not on answering predefined questions but modeling dynamic mental states and transitions using symbolic and experiential structures drawn from Buddhist contemplative psychology.

We have shown that models like GPT-4o can express nuanced, archetypally structured insight, while Mistral offers lightweight interpretive capacity for embedded systems. GPT-3.5 provides a balanced midpoint with practical use cases in reflective applications.

While we do not claim immediate readiness for mentoring or therapeutic deployment, the system’s ability to generate grounded reflections on mental dynamics points toward a broad field of future applications—from spiritual practice and self-reflection to sociopolitical modeling and strategic guidance.

Related Work

The integration of Buddhist models of mind with artificial intelligence places this project at the intersection of several emergent fields: neuro-symbolic AI, contemplative science, affective computing, and AI ethics.

Neuro-Symbolic AI approaches increasingly seek to combine symbolic reasoning with vector-based representations. This system contributes to that trajectory by embedding Buddhist psychological concepts—such as karmic causality, fear, and liberation—as high-dimensional vectors. Rather than hand-coding a formal ontology, the framework lets the meaning emerge from scriptural texts through large language model (LLM) interpretation. This mirrors the goals of knowledge infusion in modern LLMs, but with a unique source: Buddhist wisdom literature (cf. Bosselut et al., 2019).

In contemplative science, computational modeling of first-person experience has often relied on clinical or cognitive frameworks. Here, we offer an alternative: a system that draws directly from introspective descriptions in Buddhist texts. This complements empirical contemplative research by providing a symbolic-cognitive mirror to first-person phenomenology (Varela et al., 1991).

In affective computing, traditional systems recognize emotions such as joy, fear, or anger. This work extends affect modeling to include complex, process-oriented mental states such as recognition of emptiness, emergence of insight, or archetypal rescue. These transitions, described in Buddhist teachings, offer a rich map of mind dynamics for AI to interpret or respond to (Picard, 1997).

In AI ethics, the project resonates with recent calls for value-aligned AI. Buddhist ethics—emphasizing compassion, clarity, and transformation—inform not only the content but the mode of interaction between human and machine. This echoes work by Hershock (2012), Singler (2019), and others who explore spiritual foundations for humane technology.

Only a handful of projects have explored Buddhist psychology as an operational cognitive model in AI. Lou (2017) attempted a rule-based system based on Abhidhamma; this work instead lets a language model read Buddhist texts, reflecting a move from programmed to emergent models of mind. In doing so, it opens the door to a new genre of reflective, contemplative, and meaning-aware artificial intelligence.

Conclusion

This paper introduced a proof-of-concept framework for modeling mental state dynamics using Retrieval-Augmented Generation (RAG) grounded in Buddhist contemplative literature. By leveraging Buddhist symbolic and psychological insights—particularly from the Bardo Thödol—we constructed a system that interprets mental state transitions with nuance and resonance.

Comparative results from Mistral, GPT-3.5, and GPT-4o demonstrate that large language models can meaningfully reflect both the cognitive and symbolic layers of these teachings, with GPT-4o offering especially rich interpretive potential. Our work opens new pathways for applying AI to the modeling of inner experience, not merely for knowledge extraction but for engagement with mind as a dynamic, evolving field.

Future developments will expand the corpus and refine the system’s ability to interpret complex mental transitions, paving the way for applications in contemplative tools, AI mentoring systems, and reflective human-AI dialogue. With further validation and collaborative input, we believe this approach can support a new class of cognitive technologies: ones that resonate with the human journey toward clarity, transformation, and ethical insight.

References

Bosselut, A., Rashkin, H., Sap, M., Malaviya, C., Celikyilmaz, A., & Choi, Y. (2019). COMET: Commonsense Transformers for Knowledge Graph Construction. Proceedings of ACL 2019.

Hershock, P. D. (2012). Valuing Diversity: Buddhist Reflection on Realizing a More Equitable Global Future. SUNY Press.

Lou, Y. (2017). Is Buddhist Abhidhamma psychology programmable? SSRN Working Paper. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3077596

Picard, R. W. (1997). Affective Computing. MIT Press.

Singler, B. (2019). The AI Creation Meme: A Case Study of the New Visibility of Religion in Artificial Intelligence Discourse. Religion and Digital Media, 25–38.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press.