The ability to learn and adapt to cognitive demands is vital for survival, and mice serve as valuable models in neuroscience for studying these processes. In this study, we introduce a novel maze framework to investigate learning and adaptability in mice, specifically focusing on humanized App knock-in models of Alzheimer's disease. The maze mimics decision-making scenarios, similar to when foraging, with an optimal reward path leading to a target with sucrose pellets. Decision nodes introduce off-reward routes to dead-ends or loops, creating a challenge requiring mice to balance exploration with goal-oriented behavior. The study involved a 13-hour overnight trial using both Wild-type (WT) and AppSAA knock-in Alzheimer’s mice. Over time, mice displayed increased preference for the target zone, but AppSAA mice exhibited significant impairments, including spending less time at the target, more frequent deviations to off-reward paths, and taking longer routes to the target. We developed a hierarchical probabilistic framework to analyze spatial learning and decision-making in the maze, influenced by non-local cues and reward anticipation. The first level employs a discrete-time Hidden Markov Model (HMM) to capture navigational states, reflecting motor actions and movement strategies. The second level combines a Bayesian Gaussian Mixture Model (BGMM) and a Gaussian Mixture Model HMM (GMM-HMM), incorporating spatial data and inferred patterns to capture the temporal evolution of reward-oriented strategies. This approach provides a detailed view of fine-scale behavioral dynamics and the development of goal-directed navigation, revealing the cognitive mechanisms driving decision-making. The maze was enhanced with timed rewards and simultaneous electrophysiological recordings, enabling multi-modal analysis linking behavior to neural activity. This framework uncovers how specific brain dynamics support cognitive flexibility and learning, offering insights into the neural basis of decision-making during complex navigation.