- Abstract: Lifelong and meta learning aim to learn quickly new skills by transferring knowledge from past experiences. Current algorithms learn a meta-policy from a set of hand-crafted training tasks and adapt their policy to new but similar problems. In contrast, human are able to learn from very sparse data and apply learned skills to very different situations. In this work, we integrate the old-standing idea of knowledge-based agents with deep model-based reinforcement learning. We show that our model can transfer knowledge from completely randomly generated environments to real tasks. We argue that the weights of forward models can be used as a knowledge-base by an inference engine to infer optimal policy in real environments. We demonstrate that this approach automatizes both task-specification and knowledge base construction.