Voyager divides cumulative learning across distinct modules: a curriculum proposes tasks, while the skill library stores successful executable skills and retrieves them later. Generated programs compose retrieved skills, and iterative prompting uses environment feedback and execution errors to refine each program before successful skills are retained. This makes Voyager relevant to Open Epistemic Intelligence as reusable artifact memory, without treating it as evidence of decentralized ownership or sovereign learning.
An agent playing Minecraft with no human in the loop and no model weights changing. Three parts carry it.
A curriculum proposes the next task from what the agent can currently do. A skill library stores working solutions as executable programs, indexed so they can be retrieved later. A prompting loop feeds execution errors and environment feedback back in until the program runs.
Skills compose, so later tasks call earlier ones and the agent gets further into the tech tree than agents that start fresh each episode. The library is the memory, and it is readable code rather than a hidden state.