Browser automation agents forget everything between runs. Same site, same page structure to figure out, every single time. Wasted time, wasted tokens.
site-memory lets your agent remember sites it has visited and go straight to the goal.
On the WebVoyager benchmark, it cuts cost by 70% to 90% and makes tasks 4x faster with no loss in accuracy.
- Learns site knowledge automatically
- Updates after every visit
- Gets better over time
Without site-memory, booking a hotel on Booking.com takes 35 steps and nearly 8 minutes. With site-memory, the second run takes 5 steps and 30 seconds. Stable sites get even faster after a few visits.
Let your AI install it. Tell Claude Code / Codex / OpenClaw / Gemini CLI:
Install the site-memory skill from github.com/LittleYier/site-memory
Or use the Skills CLI:
npx skills add LittleYier/site-memoryOr clone manually:
git clone --recurse-submodules https://github.com/LittleYier/site-memory.gitThen copy skills/site-memory/ to your agent's skills directory
(e.g. ~/.claude/skills/, ~/.gemini/skills/, or .agents/skills/).
- Node.js 22+
Works with any browser tool your agent already has:
| Tool | Notes |
|---|---|
| Chrome DevTools Protocol (bundled) | Installed with site-memory, requires Chrome with remote debugging |
| browser-use | npx skills add anthropics/browser-use |
| Playwright MCP | Add the MCP server to your agent config |
| Claude in Chrome | Browser extension, accessed as an MCP tool |
| Others | Anything that can navigate, read, and interact with pages |
Most agent runtimes (Claude Code, Gemini CLI, Codex) can load multiple skills at once. site-memory works with whatever browser tool you pick.
Best for repeated browser work, such as:
Fill out the insurance quote form on example-insurance.com
Search for apartments under $300k on realestate-site.com
Go to the admin dashboard and export last month's report
Or call it directly:
/site-memory open books.toscrape.com and get page 3 titles
~/.site-memory/
Tested on WebVoyager tasks. 15 websites, 50 tasks, 3 rounds each (150 total runs). All correct.
Round 1: Claude Sonnet 4.6. Rounds 2 and 3: Claude Haiku 4.5 (3x cheaper).
| Site | R1 avg cmds | R2 avg cmds | R3 avg cmds | R1 avg time | R3 avg time | Cmd delta | Time delta |
|---|---|---|---|---|---|---|---|
| Booking | 35.7 | 6.7 | 5.0 | 497s | 30s | -86% | -94% |
| Amazon | 14.7 | 2.0 | 2.0 | 183s | 12s | -86% | -93% |
| ArXiv | 9.5 | 1.5 | 1.5 | 143s | 9s | -84% | -94% |
| Google Flights | 23.3 | 10.3 | 4.0 | 173s | 47s | -83% | -73% |
| ESPN | 8.3 | 2.3 | 2.0 | 85s | 18s | -76% | -78% |
| Huggingface | 17.7 | 4.3 | 4.3 | 360s | 43s | -75% | -88% |
| BBC News | 7.7 | 3.0 | 2.0 | 115s | 22s | -74% | -81% |
| Allrecipes | 3.8 | 1.0 | 1.0 | 85s | 3s | -73% | -96% |
| Coursera | 3.0 | 2.0 | 1.0 | 45s | 5s | -67% | -89% |
| Google Map | 10.7 | 5.3 | 7.0 | 273s | 159s | -34% | -42% |
| Google Search | 7.5 | 4.0 | 5.3 | 48s | 38s | -29% | -21% |
| Cambridge Dictionary | 3.0 | 2.3 | 2.3 | 48s | 30s | -23% | -37% |
| Wolfram Alpha | 4.8 | 5.0 | 3.5 | 45s | 21s | -27% | -53% |
| GitHub | 2.3 | 2.3 | 2.0 | 33s | 16s | -13% | -52% |
| Apple | 4.7 | 2.3 | 6.7 | 60s | 72s | +43% | +19% |
| Metric | Round 1 | Round 2 | Round 3 | R1 → R3 |
|---|---|---|---|---|
| Total commands | 498 | 177 | 162 | -67% |
| Total time | 6,944s | 1,746s | 1,674s | -76% |
| Avg commands / task | 10.0 | 3.5 | 3.2 | -68% |
| Avg time / task | 139s | 35s | 33s | -76% |
After just one visit, the second run is already 3x faster. For stable sites (Allrecipes, Coursera, ArXiv), the agent goes straight to the goal after a few visits.
MIT