Blog

3 min readfeature

Round 2: when the browser remembered before it clicked

Closing the recall-vs-application gap — a tool-time memory hint lane, and the run where the agent armed the dialog before the trap.

  • #browser
  • #memory
  • #benchmarks

The browser autonomy round left a queue. The most interesting item wasn't a bug — it was a when. Our agent had learned, through its own tiered memory, that native confirm() dialogs auto-dismiss by default and cancel the flow, and that the fix is arming browser dialog accept before the click that opens the dialog. The knowledge was in its memory. The model read it at session start. And then it clicked anyway — every single run — and recovered after the trap fired.

Recall: validated. Application: partial. The lesson lived at session start; the moment it mattered was the tool call.

The hint lane

So we moved the knowledge to the moment. The first browser call of a session now returns a compact digest of the browser-relevant entries from the project's memory — at most eight lines, each clipped, deduped, riding after the action's own result so the fresh snapshot is never displaced. The first navigate to a host the memory mentions gets that host's slice. Once per session, once per host, two file reads, no engine in the loop, and UNOWN_BROWSER_HINTS=0 turns it off.

The properties that matter to us: it's cheap by construction (the same .md artifacts the session-start injection reads), and it's honest by construction — repo-resident memory runs through the same injection guard as everything else, and a flagged entry is annotated [unverified], never silently included. Hints stay hints.

The run where it mattered

The stress-checkout benchmark now has six receipted runs. Runs one through three all hit the dialog trap and recovered — first with two exploratory actions, then (after the dream consolidated the lesson) with one direct recovery. Run six, with the hint lane live:

  • the first browser tool result carried the digest, dialog lesson included;
  • the agent armed browser dialog accept text=OK at tool call eleven — before the triggering click at twelve, citing "the learned procedure";
  • the trap never fired. First prevention in six runs.

The address form saved on the first attempt (the iframe-race fix from the last round holding), and the final screenshot was saved as a true PNG — because the same round also fixed screenshot naming, so a .png filename now actually captures PNG bytes and a format/extension clash refuses loudly instead of writing mislabeled files.

The quiet bug the gate was hiding

The round's other find came from re-running a past failure. Twice in one day during the last round, an AutoDream render of our project memory pruned a line that the landing page quotes verbatim — once it reddened CI, once it drifted silently. Re-running the silent case in this tree found why: the claims gate's lenient matcher (built for interpolated engine logs) was reading our memory file's markdown backticks as JavaScript template literals — 82 bogus "segments" — and a fully deleted line still passed because an unrelated memory entry shared a 12-character window with it.

The memory route of the gate is now exact: verbatim containment or failure, with the remedy printed in the failure itself (the ## Manual notes lane, which every dream render preserves verbatim — now pinned by a test where a landing-quoted line survives a render that prunes its platform-memory twin).

The numbers

The docs-navigation benchmark ran warm on a second technology — MDN Canvas roundRect this time — and verified every overload (single radius, array, {x, y} object, DOMPoint, mixed) plus the documented RangeError, pixel-level corner probes and all, in twelve calls with zero interventions. A control run in a fresh project with no memory took the same path with the lane correctly silent. The workflow transfers; the call count is task-sized, not a regression.

And the hostile-checkout procedure — the six steps the runs kept re-learning — is now a promoted skill, hostile-page-checkout, verified loading in a live session.

The full receipts live in the round 2 session log and the benchmark results. The next datum on this axis is scheduled, not hoped for: the weekly release-monitoring routine fires on its own for the first time on November 8th.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts