We introduce a generalisation of smooth fictitious play with bounded m-memory strategies. We use this learning algorithm to prove a Folk theorem from learning in repeated potential games. If a payoff profile is supported by an m-memory pure strategy subgame perfect equilibrium, then there is a non-zero probability of learning an m-memory strategy profile that is arbitrarily close to the desired equilibrium, so that the desired payoff profile is (approximately) achieved in an appropriate continuation game. Our results prove that autonomous artificial agents who use a learning process to develop strategies to play the game can learn to collude.