This thesis addresses the stochastic modeling of action potential dynamics generated by mouse retinal ganglion cells in response to specific visual stimuli. The primary objective was to characterize the photoreceptor response, evaluate the validity of various stochastic models, and compare observed firing patterns with the proposed simulations. The research began with the hypothesis that the photoreceptor response could be described by a memoryless process. Statistical goodness-of-fit tests consistently rejected this hypothesis, revealing that retinal dynamics possess intrinsic memory properties. Given this finding, a second hypothesis was evaluated, proposing that the system could be approximated as a memoryless process through the selective censoring of the 5% of extreme timing values. Despite visual improvements in the histograms, repeated statistical tests also refuted this hypothesis, confirming the persistence of memory within the system even after the removal of outliers. The rejection of memoryless hypotheses led to the development and evaluation of a memory-based model prototype, grounded in a Langevin-type model with memory that incorporates the photoelectric effect on the retina. Simulation and analysis of the results demonstrated that the model successfully reproduced the general trend of neuronal activity across various cells and different stimulation angles. Analysis of variances and standard deviations confirmed the model's robustness for most angles, although deviations were observed in highly active cells, suggesting potential limitations in temporal precision under conditions of dense activity. Methodologically, this thesis contributes to stochastic modeling with memory, develops computational tools in Python, and suggests practical applications for visual prostheses, while acknowledging limitations regarding generalizability and fine-grained precision.