Replay-Consistent Effect Models: Toward a Goal-Free Foundation for Agent Learning

We introduce a novel learning paradigm for agents based on replay-consistent predictive effects rather than reward optimization. Departing from the traditional reinforcement learning framework, our approach models each action as a locally learned transformation (Δ) of state, accepted only if statistically stable and reproducible through prior experiences. By leveraging variance-gated updates and context switching upon prediction failure, the agent autonomously adapts to changes in environmental dynamics without any external reward signal. This method lays a foundation for a goal-free theory of agent learning, where behavior emerges through stable and interpretable cause-effect relations, offering a scalable and intrinsically robust alternative to reinforcement learning.