- Abstract: Deep lifelong learning systems need to efficiently manage resources to scale to large numbers of experiences and non-stationary goals. In this paper, we explore the relationship between lossy compression and the resource constrained lifelong learning problem of function transferability. We demonstrate that lossy episodic experience storage can enable efficient function transferability between different architectures and algorithms at a fraction of the storage cost of lossless storage. This is achieved by introducing a generative knowledge distillation strategy that does not store any full training examples. As an important extension of this idea, we show that lossy recollections stabilize deep networks much better than lossless sampling in resource constrained settings of lifelong learning while avoiding catastrophic forgetting. For this setting, we propose a novel dual purpose recollection buffer used to both stabilize the recollection generator itself and an accompanying reasoning model.