Site memory: a new task on a known site, 33% cheaper
I build Dejo, so read this with that in mind. The numbers come from a benchmark run on 10 November 2026; the raw data is linked at the end, and I say where it didn't help.
Short answer: a browser agent that remembers a website gets cheaper even on a task it has never done there. We gave Claude Code a new task on 10 real websites, once with an empty memory and once after Dejo had learned a different task on the same site. With the site already known, the new task averaged 238K tokens instead of 354K, 33% fewer, and every answer was still right. On 3 sites it made almost no difference, and on one it cost more.
The test
Each site had two tasks from our real-site benchmark, call them A and B. For each site we ran task B two ways, each in a brand-new Claude Code session (Claude Sonnet 5, medium effort, the Dejo code released as 0.5.1):
- cold: Dejo's memory is empty.
- warm: Dejo first learned task A on that site in another session; then task B runs in a new one.
Task B is never the same as task A. On scrapethissite.com, for example, A was "find the hockey team stats for the Mighty Ducks of Anaheim" and B was "go to page 5 of the hockey team table and tell me the first team listed".
The results
| site | task B, cold | task B, warm | change |
|---|---|---|---|
| rfc-editor.org | 379K | 135K | -64% |
| doververmont.com | 1,156K | 537K | -54% |
| demowebshop.tricentis.com | 224K | 132K | -41% |
| collection.internationalfolkart.org | 324K | 227K | -30% |
| catalog.ccv.edu | 414K | 301K | -27% |
| packages.debian.org | 224K | 178K | -21% |
| practicesoftwaretesting.com | 234K | 230K | -2% |
| scrapethissite.com | 174K | 175K | 0% |
| tcatbus.com | 226K | 228K | +1% |
| pub.dev | 183K | 234K | +28% |
| average, 10 sites | 354K | 238K | -33% |
All 20 answers were right. An 11th site, talbothistory.pastperfectonline.com, is left out of the average because both runs gave the wrong answer there.
Why it helps
The biggest wins came where task B needed pages task A had already walked through. On doververmont.com, the town's site, task A found the highway department and task B needed the transfer station's prices. Cold, the agent took 20 turns to find its way. Warm, it took 11, with the map of the pages task A had already walked.
What Dejo keeps about a site, besides each saved task:
- The map: every page it visited and what was on it. Your agent asks for it with
dejo_memory. - Site facts: which pop-ups and cookie banners appear and how they close, so the next run clears them without a model call.
Where it didn't help
- Small sites with short tasks. On scrapethissite.com and tcatbus.com, task B took 4 or 5 turns either way. There was nothing to skip.
- pub.dev cost more. The warm run made one extra turn (5 instead of 4) and used 234K tokens instead of 183K. With one run each way, I can't tell a real cost from noise there; it is in the table as it came out.
How this fits with replay
This is the middle case. The cheapest case is a task Dejo already knows: one dejo_run call, about 89K tokens on average in our benchmark, with 17 of 19 tasks done in a single call. The most expensive is a site Dejo has never seen. A new task on a known site sits in between, and the more of the site your agent has used, the closer it gets to the cheap end. I wrote about the first two cases in browser agent memory and why Playwright MCP uses so many tokens.
How we measured
- 11 sites, each with two tasks from the real-site benchmark; each task B run once cold and once warm, every run in a brand-new session.
- Claude Sonnet 5 at medium effort, Dejo at commit
d78208a(released as 0.5.1), 10 November 2026. - Raw data:
benchmark-site-memory-2026-11-10.json, made bybenchmark_site_memory.pyin the same folder.
One run per site per way is a small sample. Read the per-site numbers as examples and the average as the signal.
Related reading: browser agent memory, browser agent token statistics and why Playwright MCP uses so many tokens.
FAQ
Does a browser agent get cheaper on a site it already knows?
With Dejo, usually yes. We gave a fresh Claude Code session a new task on 10 real websites, once with an empty memory and once after Dejo had learned a different task on the same site. The second way averaged 238K tokens instead of 354K, 33% fewer, and every answer was still right.
What does Dejo remember about a site besides a task?
The pages it visited and what was on them, plus site facts such as which pop-ups appear and how they close. On a new task your agent can ask for that map with dejo_memory instead of exploring from the home page.
Is it cheaper on every site?
No. On 3 of the 10 sites it made almost no difference, and on pub.dev the run with memory cost more (234K vs 183K tokens). The biggest savings came on sites where the new task needed pages the first task had already found: up to 64% fewer tokens on rfc-editor.org.
Is this the same as replaying a saved task?
No. A replay runs a task Dejo already knows in one call, about 89K tokens in our benchmark. This test is about a task Dejo has never done, on a site it has seen. The agent still has to work the new task out; it just starts with a map.