Metrics center

Retention

Retention is not one grade. In Fluorgan, it is a reading of memory through two lenses: predicted estimates the chance of remembering now; observed shows what happened in recent qualified reviews.

Two lenses of memory

Predicted anticipates; observed checks. One does not replace the other because they answer different questions.

Memory changes over time

Even without reviewing, prediction can fall because time passed. That does not create an observed error; it creates a lower current estimate.

Predicted vs observed

Divergence can suggest an optimistic model, small sample, recent difficulty, or a real change in the material. A prudent reading preserves that uncertainty.

Explore an example

An educational example with 12 cards shows memory state, time passing, valid review, and recovery without inflating the numerator.

What counts as a valid review

The first primary, due attempt in a consolidated phase enters the contract. Learning, same-cycle recovery, not-due Safe Ahead, variant, pretest, contrast, and probes stay out of this metric without losing operational value.

How we calculate

Current predicted retention consumes the public contract formula; observed consumes successes over qualified reviews; Retention Over Time uses historical prediction persisted before the answer.

Evidence changes the reading

The same raw percentage can carry different weight. Missing evidence is not zero, and a technical error is not empty.

How to interpret combinations

High/low in this block is only a synthetic teaching scenario. The page creates no new thresholds.

What Retention does not measure

Retention does not measure content that never became cards, complete understanding, transfer to a new context, exam performance, approval, or absolute card quality.

How it appears in Fluorgan

The simplified representation shows where value, predicted/observed lens, trend, evidence, and interpretation path appear in the Progress surface.

Contract and review

Public details show contract version, policy, window, timezone, states, population, and last review without exposing SQL or unnecessary internal names.