Abstract

The growing demand for personalized and reliable intelligent tutoring systems highlights the limitations of conventional large language models in cognitive adaptation, pedagogical reasoning, and factual consistency. This research proposes an adaptive neuro-symbolic multi-agent framework integrating self-reflective retrieval-augmented generation (RAG) with cognitive knowledge tracing to enhance personalized learning. The proposed architecture coordinates specialized agents for learner-state estimation, knowledge retrieval, pedagogical reasoning, adaptive assessment, and feedback generation. A neuro-symbolic reasoning mechanism combines neural representations with structured educational knowledge to support interpretable instructional decisions. Self-reflective RAG enables evidence verification and iterative response refinement, while cognitive knowledge tracing dynamically estimates student mastery and identifies learning deficiencies. The framework is designed to optimize instructional pathways, reduce unsupported responses, and improve assessment relevance through continuous learner feedback. The proposed system will be evaluated against conventional LLM-based, RAG-based, and multi-agent tutoring approaches using factual accuracy, knowledge-tracing performance, learning gains, pedagogical relevance, and computational efficiency. This research aims to establish a scalable, explainable, and cognitively adaptive architecture for next-generation intelligent educational environments.