Should an AI agent remember how a website works?
An agent should remember how a site is used, not what it said last Tuesday. Search boxes, working buttons and last failures are shape. Copy, prices and personal data are content. Shape is a private hint for the next visit. Content is better fetched live. Notes never override your rules or the page as it is now.
What is the difference between a site’s shape and its content?
Shape is the map of use: where search lives, which control submitted a form, which path timed out, which login wall appeared. It is structure and sequence, not the words on the screen. Content is the payload: article text, product copy, account figures, comments, and anything that belongs to a person or a publisher.
Keeping those apart is a privacy and accuracy habit, not a branding choice. Shape can sit in a short note you could read back. Content should be read from the live page, because it changes and because storing it is rarely needed for the next click. A simpler bookmark or a handwritten checklist is the better tool when you only need one URL and no agent at all.
Why is remembering shape useful, and when is a simpler tool better?
A remembered shape saves the same dead ends. If last time the filter sat behind a menu, the next visit can try that path first. If a button labelled Submit did nothing, the note can say so. That is operational memory: how the machine of the site behaved for you, not a copy of the machine’s output.
It is still only a hint. Sites redesign. Cookies expire. A/B tests swap labels. The live page always wins. If your work is a weekly visit to one form, a password manager and a browser shortcut beat any agent. If the job is many tools in one thread, a single conversational surface with memory between sessions can keep the map without you repeating it. GROX is one such surface; it is not required for a sticky note.
| Kind of memory | What it holds | Safer default | When to skip it |
|---|---|---|---|
| Shape | Paths, controls, failures | Keep a short, readable note | One familiar form |
| Content | Text, prices, names | Fetch the page again | Anything personal |
| Session crumbs | Open tabs, half-filled fields | Clear when the job ends | Public kiosks |
| Publisher copy | Articles and listings | Link, do not archive wholesale | Fair-use reading only |
Why should page content stay out of long-term memory?
Page text is someone else’s work and often someone else’s data. Storing it “just in case” creates a second copy you did not need. It also goes stale: a remembered price is not a price. An agent that treats yesterday’s HTML as truth will act on fiction. Fetch, use, discard is the honest pattern for content.
Personal fields are worse. Names in a comment, an invoice line, a message preview — none of that is “how the website works”. If a task needs those values, they belong in the current session, under your eyes, then gone. Trade-off: you type a bit more next time. Benefit: you do not build a private archive of other people’s pages.
- Shape
- How controls and paths behaved last time, written as a hint you can edit.
- Content
- What the page displayed: copy, figures, media, personal fields.
- Hint
- A note the agent may try first, then check against the live site.
- Instruction
- A rule you set that must not be quietly overwritten by a note.
Are remembered notes instructions, and how should they fail?
Notes are hints, never instructions. An instruction is a limit you chose: do not send, do not pay, stop. A hint is “search was top right last month”. If the hint is wrong, the agent looks again and says the path changed. If the hint is treated as law, you get clicks on ghosts.
Write notes in plain words you can correct. Prefer “filter failed twice” over a script of selectors that will rot. When a request is vague, a short card of questions before work starts is kinder than a confident tour of an outdated map. GROX can persist memory between sessions and route work to tools; that does not make a site-shape note into a command. Pause, change or stop with a sentence whenever real actions are in play. A practice run with nothing at risk is the right first pass for anything that spends.
- Keep shape notes short and readable.
- Re-read the live page before acting.
- Treat failures as facts about last time, not forever.
- Use a bookmark when the map is one URL.
Common questions
Should an AI agent store the text of every page it opens?
No. Page text is content, not shape. It goes stale, it may be personal or copyrighted, and it is not required to find the search box next time. Open the page again, use what you need in the session, and leave the archive to the site that published it.
Is it useful to remember which button worked last time?
Yes, as a hint. A note that the working control was in a menu can skip a dead end. The live page still decides. If the label moved, say so and follow what is on screen now, not the note. A simpler checklist is enough when you only use one form.
What should happen if a remembered path fails?
Treat the note as wrong for this visit. Look at the page, try the visible controls, and update the hint in plain words. Do not loop the old path. Do not invent a workaround that submits data you did not ask to send. Failure of a hint is expected when sites change.
When should I not use an agent for a website at all?
When the job is one URL, one login, and a form you already know. A browser, a password manager and a written step list are simpler and easier to audit. Use an agent when many tools must keep one thread, not because memory sounds impressive.