This dissertation, based on joint work, explores the role of memory in shaping belief formation of financial market participants. We estimate a structural machine learning model of memory-based belief formation applied to consensus earnings forecasts of sell-side stock analysts. The estimated model reveals significant recall distortions compared to a benchmark model trained to fit realized earnings revisions. Specifically, analysts over-recall distant historical episodes most of the time, when recent events are more useful for forming forecasts than those in the distant past, but under-recall them during crisis times, when history helps to interpret unusual events. We document two potential driving forces behind these distortions. First, analyst memory overweights the importance of past earnings and forecasts. Second, analysts are more likely to selectively forget past positive events. Our model of analyst recalls strongly predicts their earnings forecast revisions and errors, as well as stock returns, which suggests that distorted recalls might contribute to mispricing of assets in financial markets