An agent that must remember its actions on a web page needs a criterion for when two pages count as the same. Existing memory schemes based on observation similarity merge pages that look alike but behave differently, which is common in GUIs: two tabs of the same widget or two rows of a menu may respond to the same click in distinct ways. We define the merge rule as an action‑conditioned bisimulation applied to the empirical predictive state graph that a frozen agent populates while acting. Two states are merged only if their shared actions, under the same affordance label, lead to agreeing outcomes and to successor blocks that belong to the same equivalence class. Observation similarity never enters the rule, and no training is required. This rule replaces the merge component of an existing outcome‑value memory, allowing a closed‑loop comparison to isolate its effect. In MiniWoB++ experiments the agent using action‑conditioned bisimulation achieves a higher success rate than a memory‑less baseline, while a control that follows identical exploratory detours, the prior successor‑representation merge, and the same criterion without action conditioning produce no improvement.
Review: The study demonstrates that a strict action‑conditioned equivalence relation can improve GUI‑based memory merging without additional training, offering a promising direction for more robust web‑interaction agents in reinforcement learning.