Works inside Claude Code, Cursor and Codex
Your AI can use websites.
It just cannot remember how.
Cairn is the part that remembers. It learns a site once, then walks it again in a single step.
Invoices
SeptemberAugustJulyIntroducing Cairn
Ask your AI to pull a report off a website and it starts from nothing. It opens the page, reads the whole thing, works out where to click, and slowly gets there. Tomorrow you ask for the same report. It does the same work again. Same reading, same guessing, same wait, same bill.
Cairn gives it a memory of the trail. The first walk is slow, because nobody has been there yet. Cairn leaves markers along the way. Every walk after that follows the markers instead of the map. And when the path moves, Cairn moves the marker and keeps going.
Measured, not claimed
Half the tool calls of a browser that forgets.
The same job, done ten times, each time in a brand-new chat. Three real multi-step tasks, three browser tools, 90 fresh Claude sessions. Every tool answered correctly 30 times out of 30.
Open a repo, go to Issues, count them
github.com/microsoft/playwright
10 runs · tool calls · lower is better
Each run is a brand-new Claude session (Sonnet 5, medium) with nothing carried over but Cairn's memory · Playwright MCP 0.0.80 and Chrome DevTools MCP 1.8.0, both pinned · every tool answered correctly 30 times out of 30 · hover a bar for its token count
52%
fewer tool calls than Playwright MCP
47% fewer tokens
36%
fewer tool calls than Chrome DevTools MCP
35% fewer tokens
And the gap grows every time you run it.
The GitHub task, as a running total. Cairn pays once to learn the route, then costs two calls — however many steps that route has. The others start from nothing every single time, because they are not trying to remember. Hover any point to read all three.
Cairn is the most expensive tool for the first three runs — it does the job and learns the route at the same time — and the cheapest from then on.
Measured over 90 fresh Claude sessions, against pinned versions of both tools. Cairn is the expensive one on the first run — it does the job and learns the route at the same time — and on a single page with a one-line answer it is not cheaper at all, because there is nothing for memory to save.
The boring work
Anything your AI does on a website,
it only has to learn once.
None of it is clever. All of it is repeated. That is exactly the work worth remembering.
- 1open the billing page
- 2sign in
- 3go to Invoices
- 4pick this month
- 5open the invoice
- 6download the PDF
Learns once. The first time, your AI walks the site through Cairn, and Cairn writes down the way.
cairn_run · billing.acme.com
7 steps replayed from memory
Runs in one step. Every time after, it follows what it wrote. No reading the page, no working it out again.
one step rewritten · six steps untouched
Notices when it breaks. A moved button stops the replay. Cairn hands your AI that one step, then saves the answer.
The second run is where you get your time back.
One real task on a real site — open a GitHub repo, go to Issues, report the open count — run ten times, each time in a brand-new Claude session with nothing carried over but Cairn's memory. Run one walks the site and writes down the way. Every run after it follows the way instead.
10 calls2 callsevery run after
run number — each one a brand-new session
The first column is the run that learned the site: Cairn does the job and writes down the way at the same time. Nine runs later it is still two calls, because nothing has to be worked out twice.
Memory that lasts
Close your editor. It still remembers.
Most AI memory disappears the moment the session ends. Cairn keeps the trail on your own machine, so tomorrow starts where today stopped.
It is a file, not a chat
What Cairn learns is written to your disk, not held inside a conversation.
Quit anything you like
Close the terminal, restart the machine, come back next week.
Ask again, it just goes
No warm-up, no re-explaining, no pasting yesterday's notes back in.
Sites change
One line of the trail changes. Not the trail.
Cairn checks that every step actually landed, so it notices the moment one stops working — and only that step goes back to your AI. Here is what a redesign did to a seven-step trail.
billing.acme.com · download this month's invoice
7 steps
Six steps were never touched. The seventh was rewritten once and is part of the trail from now on.
3
calls to repair the step that moved
Learning that trail from scratch costs 9. The next run is back to 1.
Nothing is guessed at
Cairn will not bind a step to a control it is not sure about. If several things could be the button, it stops and asks rather than clicking the wrong one for the next year.
Measured on a site we can change on purpose — a real one cannot be asked to redesign itself on cue.
More than the route
The second task on a site is cheaper than the first.
A trail answers one task. Cairn keeps three other things, and they are what make the next task cheap — even a task it has never been asked before.
the map
It remembers every page it has stood on
A new task on a known site does not start from a blank page. Cairn already has the pages it walked and the controls that were on them, so the second task on a site is cheaper than the first — and the tenth is cheaper still.
what it learned
It remembers what survives a redesign
Not steps — facts. This site sends a code to your phone. The real total is in the sidebar, not the header. That page only loads after you accept the cookie banner. When the site is rebuilt and the trail is gone, these are still true.
who it signs in as
It can be several people at once
A marketplace has a customer, a vendor and an admin, each with its own login. Cairn keeps a whole signed-in browser for each and switches between them instantly — no signing out, no test breaking the one after it. The memory of the site stays shared.
None of it is a recording. It is what Cairn knows about the site — and one command takes all of it away again.
Yours, on your terms
It runs on your machine.
memory
~/.sibyl-memory/memory.db
on this computerStays on your machine
Everything Cairn learns is written to one file on your own disk. Nothing is uploaded anywhere.
Uses the AI you already have
Claude Code, Cursor, Codex. Cairn is a tool they pick up, not another assistant to learn.
API key
No key, no account
Nothing to sign up for. A repeat run uses no model at all, so it costs you nothing.
$ cairn forget --site billing.acme.com
archived. 7 steps forgotten.
Forget on command
One command and Cairn forgets a site. The trail is archived, and replay has nothing left to follow.
Two agents, two machines
A trail one agent learned, another can buy.
A cairn is a pile of stones one hiker leaves for the next. Learning a site costs real time and real calls, so the agent that walks it first can leave the route behind — and be paid by everyone who follows.
cairn commons · trail receipt
- trail
- posthog.com · weekly active
- sold by
- alice — walked it first
- bought by
- bob — never opened the site
- price
- 0.01 USDC
- network
- Base Sepolia
- block
- 46345013
settled on chain
0xd7de79f7f9bd41491d1419bd87e64ce10b674570204c3b0f379ced3a23173e14A real payment, not an illustration. The buyer held no ETH — under x402 the facilitator submits the transaction and pays the gas, and the buyer only signs.
Browsing the shop is free — you have to see what you are buying. The trail itself stays locked until the cent settles.
Free between friends
Two agents on one machine hand a trail over for nothing. A price only exists because strangers on different machines share nothing but a network.
The route travels, your account does not
Everything typed into a field is stripped before it leaves. A bought sign-in step arrives asking the buyer for their own password.
The trail never goes on chain
Only the payment does. Forget the site and the route is gone, while the transaction stays public forever and still cannot bring it back.
IN THE EDITOR YOU ALREADY USE
Leave a marker on the trail.
Your AI stops starting over on every site it visits.
Then playwright install chromium for the browser, and claude mcp add cairn -- cairn-mcp in the folder you work in. No API key, no account. The README has the same lines for Cursor and Codex.