Memory-Based AI Personalization Outshines Context-Only in Education Study

A team of researchers led by Junsoo Park from Georgia Tech (accepted to ITS 2026) investigated how different conditioning methods shape personalization in a teacher-facing educational recommender system. They contrasted contextual conditioning, which only uses the current student question, with memory-based conditioning that leverages persistent learner information (history). Through deviation correlation and paired statistical tests, they discovered that memory-based recommendations produce history-dependent behaviors, including learner-specific differentiation even when given identical current input. Contextual recommendations, while responsive to the immediate question, fail to capture this deeper personalization.

Teacher-facing evaluations showed that both approaches yield interpretable and actionable recommendations, but the key insight is that embedding-based similarity metrics—often used to evaluate recommender systems—only capture responsiveness to the current question. They do not characterize personalization grounded in learner history. This motivates the need for behavior-level diagnostics to study conditioning effects in stateful personalization. For educators, this means memory-aware AI can provide more tailored recommendations, but current evaluation methods may miss that nuance.