This paper reports a cognitive psychology experiment and a Markov decision process (MDP) model of the production effect—higher memory retrieval that follows speaking aloud or writing/typing words, as opposed to lower memory retrieval when words are read silently. Current models of the production effect draw on the global-matching framework of memory. We identify four limitations of these models and present a MDP model (a perceptual active inference model) to causally explain a superior production effect of speaking over writing. University students performed a word-production task comprising speaking and writing conditions, followed by a memory test. The results showed main effects of condition on accuracy and response times. The MDP model indicated higher sensory precision during memory retrieval in the speaking condition than in the writing condition. Through Bayesian model selection, we evaluated whether the MDP model, as a mechanistic active-inference model, provided higher construct validity than a descriptive linear model (fit via Variational Laplace). The MDP model outperformed the linear model, suggesting that production modalities are hidden states that cause the visual sensory observation of words that had been linguistically produced. Crucially, the MDP model explains both group effects and individual variability, confirming the reliability paradox of statistical models.