Robust sequence prediction is an essential component of an intelligent agent acting in a dynamic world. We consider the case of near-future event prediction by an online learning agent operating in a non-stationary environment. The challenge for a learning agent under these conditions is to exploit the relevant experience from a limited environmental event history while preserving flexibility.We propose a novel time/space efficient method for learning temporal sequences and making short-term predictions. Our method operates on-line, requires few exemplars, and adapts easily and quickly to changes in the underlying stochastic world model. Using a short-term memory of recent observations, the method maintains a dynamic space of candidate hypotheses in which the growth of the space is systematically and dynamically pruned using an entropy measure over the observed predictive quality of each candidate hypothesis.The method compares well against Markov-chain predictions, and adapts faster than learned Markov-chain models to changes in the underlying distribution of events. We demonstrate the method using both synthetic data and empirical experience from a game-playing scenario with human opponents.