Remembering an entire successful browser session can preserve details that mislead the next task. The useful unit is often a smaller procedure with variable inputs: finding a product, for example, rather than repeating one particular purchase.
Key Findings
- On WebArena, reported task success rises from BrowserGym's 23.5% to 35.5%, a 51.1% relative improvement; most tasks still remain unsolved.
- On Mind2Web's cross-task split with GPT-4, step success rises from 36.2% to 45.1%. This is a step-level measure, not full-task completion.
- Online memory can develop from prior attempts; offline memory instead uses annotated examples. This makes experience availability an explicit design choice.
Who Should Read It
Useful for researchers designing persistent browser agents and comparing reusable procedures with retrieved demonstrations. Online learning needs a judge to identify successful attempts; incorrect judgments can contaminate memory. Results on controlled web benchmarks do not establish reliability across arbitrary live websites.