Spaced Repetition With AI: Making Practice Stick

Spaced repetition is a simple idea: review what you are learning at growing intervals, timed so you see each item again before it slips away. The research behind it is solid, software handles the scheduling, and AI can now take over the tedious part, writing the cards. What still decides whether it works is card quality and whether you show up every day.

What the spacing research shows

Spreading study out over time beats cramming the same amount into one session, and the research also says something about how far apart the sessions should be.

A 2006 meta-analysis in Psychological Bulletin by Nicholas Cepeda and colleagues, "Distributed Practice in Verbal Recall Tasks", pulled together 839 assessments of distributed practice from 317 experiments. A central finding: the gap between study sessions that produced the best retention grew as the retention interval (the time until the final test) grew.

A follow-up study, "Spacing Effects in Learning: A Temporal Ridgeline of Optimal Retention", published in Psychological Science in 2008, tested this over long periods. More than 1,350 people learned a set of facts, reviewed them after a gap of up to 3.5 months, and took a final test up to a year later. The best gap was about 20% of the test delay when the test was a few weeks away, falling to about 5% when the test was a year away.

Two practical readings:

• The right gap depends on how long you need to remember. Material for an exam next month wants reviews days apart. Material you need a year from now can take gaps of weeks.
• Too short and too long both cost you. In the 2008 study, lengthening the gap first helped and then slowly hurt.

One caveat: the 2008 study used a single review, while flashcard apps schedule many reviews at growing intervals. The apps build on the same principle, but the exact percentages above do not transfer directly to your deck.

How scheduling algorithms decide when you see a card

You do not have to compute any gaps yourself. The scheduler does it per card, based on how you answer.

SM-2

SM-2 comes from Piotr Wozniak, who wrote the first SuperMemo program in December 1987. As Wozniak describes the algorithm, you grade each answer from 0 to 5, and each item starts with an easiness factor (EF) of 2.5. The intervals are:

I(1) = 1 day
I(2) = 6 days
I(n) = I(n-1) * EF    for n > 2

After each review, with grade q from 0 to 5:
EF' = EF + (0.1 - (5 - q) * (0.08 + (5 - q) * 0.02))
EF never drops below 1.3

If a grade is below 3, the item restarts from the first interval without changing its EF. Easy items stretch out fast; hard items keep a lower EF and come back more often.

FSRS

The Free Spaced Repetition Scheduler, or FSRS, is a newer approach. The Anki manual presents it as an alternative to Anki's legacy SuperMemo 2 (SM-2) algorithm and, as of September 2026, lists support in Anki 23.10, AnkiMobile 23.10, AnkiWeb and AnkiDroid 2.17 or later. According to the FSRS project, it was developed by Jarrett Ye and tracks three values per card:

• Retrievability: the probability you can recall the item right now.
• Stability: the number of days it takes for retrievability to fall from 100% to 90%.
• Difficulty: how inherently hard the item is, which affects how fast stability grows after each review.

Instead of fixed multipliers, an optimizer fits the model's parameters to your own review history. You set a desired retention, the chance of remembering a card when it comes due. The manual gives a default of 90% and warns that higher targets mean shorter intervals and more reviews, with workload rising drastically as the target approaches 100%.

In practice: if your app offers FSRS and you have some review history, turn it on, run the optimizer, and leave desired retention at 90% unless you have a clear reason. Pushing it higher buys a little more recall for a lot more daily work.

Using AI to draft cards

Writing cards is where most people quit, and it is where AI helps most. It is also where AI can quietly fill your deck with junk. A workflow that avoids that:

1. Understand first. Wozniak's first rule of card writing is "Do not learn if you do not understand." Cards are for keeping knowledge, not for acquiring it. Study the material, then make cards.

2. Give the AI your source. Paste the passage you studied, not just a topic name, so the cards come from material you trust rather than from the model's memory.

3. Set the rules in the prompt. For example:

   Here is a passage I've studied and understood: [paste]
   Write flashcards from it using these rules:
   - One fact or idea per card. No lists on the answer side.
   - Each question must have exactly one correct answer.
   - Use cloze deletions for definitions and key sentences.
   - Add a short source note (section or page) to each card.
   - Skip anything trivial and anything not stated in the passage.
   Output as a two-column table: Front, Back.

4. Edit every card. Delete, merge and rewrite. Expect to cut a real share of what the model produces.

5. Add in small batches. Every new card creates future reviews, so a modest daily limit on new cards keeps the load sustainable.

Rules for card quality

Wozniak's "Twenty rules of formulating knowledge", published in 1999, is still the best checklist. The rules that matter most when reviewing AI drafts:

• Minimum information principle. Make each item as simple as possible. "What are the main features of FSRS?" is a bad card. "In FSRS, what does stability measure?" with the answer "Days until recall probability falls from 100% to 90%" is a good one.
• Avoid sets and enumerations. Asking for a full list on one card is hard to remember and hard to grade honestly. Split it, or use cloze deletions for one item at a time.
• Combat interference. Similar items get confused. A Spanish learner who mixes up librería (bookstore) and biblioteca (library) should add a card that contrasts them directly rather than two cards that sit side by side and blur.
• Personalize and provide examples. Linking an item to your own life or work makes it easier to recall.
• Provide sources. A source note lets you settle doubts later without guessing.

Also check each AI draft for three problems: a question with more than one defensible answer, a question whose wording gives the answer away, and any "fact" that is not in the source you pasted.

Review daily and keep the load honest

The scheduler assumes you review cards when they are due. Skip a few days and reviews pile up, and the intervals stop matching your memory.

• Tie reviews to a fixed daily trigger. Something like your first coffee or your commute, rather than waiting for free time.
• Keep sessions short and daily. Long, occasional sessions waste the spacing effect that the schedule is built on.
• Grade honestly. Marking a miss as a pass teaches the scheduler the wrong interval.
• Rewrite cards you keep failing. Repeated failures usually mean a badly formed card, not a bad memory. Simplify it or split it.

The number to watch is your daily review count. If it keeps climbing faster than your available time, add fewer new cards. Do not skip reviews.