Situation-Conditioned Thinking Policies for Autonomous Agents

The proposed framework replaces growing context windows with a lightweight policy that predicts required thinking patterns based on temporal or spatiotemporal situations. It consolidates long-range regularities from independent episodes into new thinking knowledge without explicit historical memory.