For AI agents: the complete documentation index is available at https://docs.flashnerd.com/llms.txt, the full documentation bundle is available at https://docs.flashnerd.com/llms-full.txt, and this page is available as Markdown at https://docs.flashnerd.com/guide/statistics/recall-and-knowledge.md.
Skip to main content

Recall, retention, and knowledge estimates

These statistics describe different aspects of learning. Retention comes from recorded answers. Retrievability is an estimated probability of recall. Desired Retention is a scheduling setting.

True Retention

True Retention is the proportion of qualifying SRS answers graded Hard, Good, or Easy. Again is a failure.

The summary uses the latest 30 study days. A review qualifies when:

  • The card was in Review at the start of the answer; or
  • It was in Learn/Relearn, and its immediately preceding review scheduled an interval of at least one day.

New-card answers and unknown phases do not qualify. Repeated qualifying answers count separately; this is a review-event statistic, not a unique-card count.

For example, if 80 of 100 qualifying reviews were successful, True Retention is 80%. The general Accuracy chart can have a different percentage because it includes a different set of answers.

Estimated Recall

The Summary's Estimated Recall is based on active studied cards with usable memory state. New, Suspended, Disabled, and cards lacking required memory data are excluded.

The model estimates each eligible card's current probability of recall and sums those probabilities into card equivalents. If two cards each have an estimated 80% chance of recall, they contribute 1.6 estimated recall card equivalents. This does not identify a specific “0.6 card” or prove either card has been forgotten.

As time passes without review, the estimate can fall even when no card has been deleted. Reviewing cards updates the model.

Recall by Time

This chart compares observed retention with estimated recall across local time-of-day groups. It uses two-hour buckets.

  1. Open Stats → SRS and choose a scope.
  2. Find Recall by Time and choose the time range.
  3. Read the All hours comparison first.
  4. Select a time-of-day bar to inspect that bucket.
  5. Check the eligible review count before interpreting a large difference.

The comparison needs both a qualifying event and usable memory data from its immediately preceding review. Some events that count toward the True Retention summary cannot be used in this model comparison.

Retention gap is observed retention minus estimated recall in percentage points. For example, 80% observed minus 60% estimated is a +20 percentage-point gap, displayed as +20%. It is not a 20% relative improvement.

The Reviewed cards label in this chart counts eligible review events, including repeated reviews of the same card. An empty bucket does not mean 0% retention; it means there were no qualifying observations for that comparison.

Time-of-day differences can reflect the material studied and the number of observations as well as your routine. A small bucket is not enough to establish a best hour to study.

Knowledge / Time

This chart reconstructs memory estimates across time using available history.

ValueMeaning
Studied cardsCards with usable memory state at that point in the reconstructed history.
Estimated recallSum of their estimated recall probabilities, expressed as card equivalents.
Average retrievabilityThe average recall probability across those cards.
Knowledge changeChange in the estimated knowledge measure over the displayed period.

The studied-card line and estimated-recall line can separate as estimated memory decays between reviews. They can move when new cards acquire usable memory history or existing cards are reviewed again.

An empty chart or No history yet is expected when the selected cards do not have usable history. Missing history is not a measured zero knowledge score.

Imported or removed history

Imported collections may not contain the same memory snapshots as reviews performed in Flashnerd. Some imported estimates are reconstructed using available intervals, so treat them as approximations rather than exact historical measurements.

Resetting review history, deleting cards, or changing which cards belong to the selected scope can change the charts. Keep a backup before destructive changes if you want to retain the underlying study record.

Related: Scheduling settings, Review history, and SRS dashboard.