Active recall sessions burn the 5-hour rolling window faster than your card count suggests

Why a recall quiz is heavier than it looks

Active recall as a study technique works by retrieval, not review. You ask Claude to quiz you, you answer, Claude grades, you move on. From the human side every turn is a card. From the API side every turn is the entire prior conversation plus one new sentence.

That asymmetry is what surprises students who run their first serious study session. Card 1 is cheap. Card 12 is meaningfully heavier. Card 27 is a small essay every time you say “next”, because the transcript Claude needs to keep grading consistently is 18 prior question/answer pairs and your running notes:

The visible turn at 16:51 is two words. The request body is whatever the running thread weighs at that point: every prompt and reply, serialized back to the model, weighed against the 5-hour cap. That is why this page exists, and why a guide that talks about active recall without mentioning the rolling cap leaves out the part that fires.

What the menu bar actually shows during the session

One row, three columns, updated every 60 seconds. The countdown is the time until the rolling window has shed enough usage that the fraction goes back under the cap; the percent is five_hour.utilization rounded for display:

The 7-day row is also there, but during a single recall session it barely moves. The signal that matters for “am I about to be cut off mid-card” is the 5-hour row. At 71 percent with 38 minutes on the clock, you have somewhere between three and seven more heavy turns before the bar trips orange, depending on how the transcript has grown. That is enough information to decide consciously whether to push or stop.

The countdown is one function, four bands

The label after the “5-hour” word comes from a single humanization function in popup.js. It bands the difference between now and resets_at into one of four shapes: “now” if the bucket is already clear, “Nm” under one hour, “Nh” up to forty eight hours, “Nd” beyond. On a study session the 5-hour row lives in “Nm” or low “Nh”, which is the band that lets you plan a break:

So “38m” is a thirty-eight minute coffee break. “1h” is a walk. “3h” is “the rest of this study session will be Sonnet, not Opus”. The function is ten lines and it is the part of the tracker that makes the percent actionable, because a percent without a clock is just an apology in advance.

The poll is fixed at 60 seconds, on purpose

The extension registers a chrome.alarms tick on install and on browser startup with a one-minute period. The same alarm fires refresh(), which calls fetchSnapshots() and posts to the menu bar app over localhost. The number is POLL_MINUTES = 1 on line 3:

A faster poll would not get you a fresher answer because the server-side bucket only recomputes that often. A slower poll would miss the moment a long Opus turn lands on the bucket, which is the moment the percent jumps from 64 to 78 in one tick during a recall session. Sixty seconds is the cadence claude.ai's own Settings page recomputes against, and matching it keeps the numbers in lockstep.

The endpoint, and why nothing else can see your study traffic

The extension calls fetch with credentials: "include" so the existing claude.ai session cookie travels with the request. The endpoint is the same internal one claude.ai/settings/usage renders against:

The reason this matters specifically for active recall sessions: most Claude usage trackers (ccusage, Claude-Code-Usage-Monitor) read ~/.claude/projects/<project>/<session>.jsonl on disk. Those files are written by the Claude Code CLI. A claude.ai chat session, which is where flashcard quizzing happens for most students, never writes to that path. Local-log tools therefore show zero usage during a recall session even when the server-side bucket is at 71 percent. The only honest read on the cap is the cookie authenticated GET, which is what the extension is doing once a minute on your behalf.

Server-truth tracker vs local-log tool, on a recall session

Both kinds of tools are useful for different things. They do not replace each other. For a study session running through claude.ai chat, only one of them sees the traffic.

Six things to keep in mind during the session

Install in four steps, before your next session

The pattern that keeps the bar green for two hours

The honest caveat

The endpoint is internal and undocumented. Anthropic can rename a field in any release. The Rust struct in src/models.rs declares each known field as Option<Window>, so a missing field deserializes cleanly and the menu bar shows an error chip instead of a wrong percent. That is a forward-compat hedge, not a guarantee. If the shape of the response moves, the open-source repo gets a same-day patch and you pull the next brew release.