Browser agent benchmark shows diff-based page memory cuts token growth 37%

Desperate_Title1595 · reddit · 2026-07-25

The author benchmarked a browser agent memory manager, Rote, against Browser Use on a live WordPress site using the same model and 150 verified runs.

What changed

Instead of re-sending the full page render every step, Rote sends only diffs after an initial grounded snapshot.
Element IDs are derived from content-based hashes so they survive re-renders and navigations.

Results

Token growth: 2,160 vs 3,437 tokens per step, a 37.2% reduction in growth rate.
849 diffs sent, with a median size of 24 characters versus 9,270-character full snapshots.
About 2.7× fewer output tokens per run.

Trade-offs

On short tasks (9 steps), Browser Use was still about 15% cheaper because its long immutable prefix benefits from provider caching.
Cost only flipped in Rote’s favor around 13 steps, reaching 16% cheaper at 25 steps.
A key limitation: eviction means the agent remembers what it did, not what it saw, so tasks that require recall from the left side of the page can fail until a notes mechanism is added.

The author also notes Browser Use ships default history compaction, which Rote does not yet have.

More from coding & agent