lifelong RL

Lifelong RL takes the long view: instead of mastering one task, an agent lives through a long, open-ended stream of tasks and is judged on how well it accumulates, organizes, and reuses skills across its whole lifetime. The picture is less a single exam and more a career — early experience should make later experience cheaper, the way a seasoned engineer solves a new problem by recombining things learned over decades.

The technical heart is building and maintaining a growing repertoire: a library of skills or options, reusable representations, and perhaps a world model that gets richer over time, all while not forgetting and while transferring forward to genuinely new situations. It therefore sits at the intersection of transfer, continual learning, and hierarchical RL, and is usually evaluated by whether the agent's learning speed on new tasks improves the longer it has lived.

Lifelong RL is more an aspiration than a solved setting. We do not yet have agents that reliably get better at learning the more tasks they see, without either forgetting or ballooning in size and compute. It is closely related to continual RL and to open-ended learning, and is one of the clearest statements of what a truly general learning agent would have to do.