Dynamic Desired Retention

Anki add-on for overriding FSRS desired retention during review with ordered Anki search rules, note-field threshold rules and answer-button grade rules.

This add-on targets Anki builds that expose gui_hooks.reviewer_will_compute_desired_retention. Grade rules additionally require gui_hooks.reviewer_will_update_scheduling_states, as do field-rule interval bounds, so they can adjust per-button scheduling states before answer button intervals are rendered.

Configure rules in Anki's add-on config:

{
  "rules": [
    {
      "query": "Frequency<1000 note:Japanese",
      "desired_retention": 0.9
    }
  ],
  "field_rules": [
    {
      "note_type": "Japanese",
      "field": "Frequency",
      "scope_query": "deck:Japanese",
      "filter_query": "tag:excluded",
      "field_value_mode": "absolute",
      "relative_top_direction": "lower",
      "value_type": "number",
      "linear_between_thresholds": false,
      "minimum_desired_retention": 0.5,
      "minimum_retention_minimum_interval": 30,
      "minimum_retention_maximum_interval": 365,
      "thresholds": [
        {
          "value": "1000",
          "desired_retention": 0.9,
          "minimum_interval": 2,
          "maximum_interval": 30
        },
        {
          "value": "20000",
          "desired_retention": 0.6
        }
      ]
    }
  ],
  "grade_rules": []
}

The first matching rule wins. If no rule matches, Anki keeps the deck/preset desired retention.

You can edit search rules, field rules and grade rules from Tools > FSRS Dynamic Desired Retention Table. In the editor, 90, 90%, and 0.90 all mean 90%. The editor displays those values as 90. When editing JSON directly, use decimal values: 0.9 means 90%. On Anki builds with RWKV-Instant integration, saving rules also invalidates RWKV's resolved targets and refreshes the current screen's due counts, so an Anki restart is not required.

Field rules can target exact text values, absolute numeric thresholds, or relative percentile thresholds such as Top 10%. Numeric field rules include a Higher DR for selector so lower field values, like low frequency, or higher field values can receive the higher desired retention. The editor keeps numeric thresholds sorted, validates desired-retention order, and always includes a Minimum Retention row for values beyond the least-important threshold. Empty numeric fields are treated as non-matches. When a numeric field contains HTML formatting, the add-on parses its visible text, so a value such as <li>17845</li> is handled as 17845. Each slice can also set optional minimum and maximum intervals in days; the add-on clamps the in-memory review scheduling states for the matched field slice. Scope Filter limits the cards a field rule can apply to, while Exclusion (Search query) skips matching cards. Conflicting threshold rows are highlighted in the editor. The threshold table also estimates each slice's workload by summing 1 / interval for matching cards with FSRS stability, using Anki's FSRS interval calculation for each card's desired retention. When Anki's batch ADR target helper reports a native Dynamic DR mapping for the active card-level preset, the estimate uses the Good-button ADR retention as the interval target. It then rounds the projected interval and caps it at the card deck's maximum interval. The editor displays workload as rounded review/day values, shows each slice's share as Workload Ratio, plus Total Workload with the same-population current workload in parentheses and workload-weighted Expected Retention below the threshold table. The current workload uses the cards' stored current intervals, which makes it easier to compare the rule scope with Anki's stats. Workload estimates run in the background; the table shows ... while they are being refreshed.

Relative percentile rules are computed from the current collection at review time and cached per field rule while the add-on is loaded, so bulk rescheduling does not repeatedly scan the same comparison population.

Grade rules are disabled by default. In the Grade Rules tab, each answer button can be left disabled, set to an absolute desired retention, or set to a relative signed percentage. Positive relative values move from the current review desired retention toward 100%; negative values move toward 0%.