The educational technology market continues to expand, yet most learning management systems remain content-centric, weakly personalized, and reactive to failure. This paper presents Lumina, a privacy-first, multi-agent adaptive learning management system that combines Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT), reinforcement learning (RL) for curriculum sequencing, retrieval-augmented generation (RAG), and a behavior engine that captures more than fifty passive learning signals. Lumina coordinates six specialized agents through the Model Context Protocol to provide closed-loop tutoring, assessment, intervention, analytics, and governance. The evaluation is organized into three complementary layers: prototype stress testing across 10,000 synthetic learner scenarios, external validation on the public xAPI-Edu-Data benchmark (480 learner records), and a reproducible benchmark matrix defined for future knowledge-tracing studies on ASSISTments and EdNet. The prototype analytics benchmark completed in 0.90 seconds with zero runtime failures and a throughput of approximately 11,051 inference runs per second. On xAPI-Edu-Data, a Random Forest baseline achieved 0.800 accuracy and 0.804 macro-F1 under five-fold stratified cross-validation, demonstrating that the behavioral signal families used by Lumina are predictive on real educational data. These results position Lumina as a technically coherent, self-hosted, and benchmark-ready foundation for explainable personalized education.

Adaptive Learning; Bayesian Knowledge Tracing; Deep Knowledge Tracing; Multi-Agent Systems; Reinforcement Learning; Retrieval-Augmented Generation; Educational Data Mining; Benchmark Evaluation