To use more IID data during SGD iterations, DQN introduced a replay memory (also called experienced replay) to collect and store the experience in a large buffer. This buffer ideally contains all the transitions that have taken place during the agent's lifetime. When doing SGD, a random mini-batch will be gathered from the experienced replay and used in the optimization procedure. Since the replay memory buffer holds varied experience, the mini-batch that's sampled from it will be diverse enough to provide independent samples. Another very important feature behind the use of an experience replay is that it enables the reusability of the data as the transitions will be sampled multiple times. This greatly increases the data efficiency of the algorithm.