Because of sorting, more skilled workers are more productive in higher type firms. Theyalso learn at different rates about their productivity and therefore have different wage paths across firms. Unlike most standard models of learning and continuous time games of incomplete monitoring, there is no deterministic outside option (absorbing state): agents always learn in all firms, albeit at different rates. This imposes a no-deviation condition which captures sequential rationality. It results in a restriction on the second derivative of the value function in addition to smooth-pasting and value-matching. We completely characterize the equilibrium, including the distribution of beliefs and the fact that under supermodularity there is always Positive Assortative Matching. Surprisingly, this holds even if the signal-to-noise ratio is higher in the low type firm. The model provides realistic predictions that existing learning models of the labor market cannot generate. For example, wage dispersion increases with tenure and an indifferent type will accept lower wages at a high type firm because of learning.