hourlong learnings #6: active recall, Anki, and retention
In the following post I try to understand why retrieving an answer changes memory more than seeing it again, why Anki is revered in the pre-med and med student community, what its FSRS scheduler is estimating, whether reviewing cards on a treadmill helps, and what exam research actually proves
active recall and spaced repetition solve different problems. Active recall makes the brain produce an answer without seeing it, while spacing decides when to demand that answer again
as usual, a basic thought map is below:
$$ \text{encode} \rightarrow \text{wait} \rightarrow \text{retrieve from a cue} \rightarrow \text{check} \rightarrow \text{schedule the next attempt} $$
Anki automates the last step
1. recall changes the task
Recognition asks whether an answer looks familiar once it is visible. Recall asks the brain to produce the answer from a partial cue
recognition
question + visible answer
→ familiarity can feel like knowledge
recall
question only
→ search memory
→ commit to an answer
→ reveal and compareRereading mostly repeats the original input, but a card with the answer hidden changes the task from taking information in to reconstructing it
That difference resembles an exam and a clinical decision more closely. Both supply incomplete cues, competing possibilities, and a time limit rather than the page where the answer appeared
what the brain is doing
Declarative memory isn’t stored as one file in one location. Features are represented across cortical networks, while the hippocampus helps bind their relationships during early learning
During retrieval, a partial cue can drive hippocampal pattern completion and reinstate parts of the cortical activity present during encoding. Human intracranial recordings found earlier reinstatement of item–context associations in the hippocampus followed by later reinstatement of item information in lateral temporal cortex
$$ \text{cue} \rightarrow \text{hippocampal pattern completion} \rightarrow \text{cortical reinstatement} \rightarrow \text{answer} $$
Retrieval also creates another learning event. The recovered memory becomes active, feedback can correct it, and its connection to the cue can be updated. An fMRI study of retrieval practice found greater anterior and posterior hippocampal activation during a delayed test than after additional study
The imaging supports a mechanism, but the behavioral conclusion is stronger: successfully retrieving information usually makes later retrieval more likely
effort needs feedback
A difficult attempt can help because the learner must reconstruct the target and discriminate it from competitors. Effort alone isn’t the treatment
too easy
→ little diagnostic information
hard but recoverable
→ effortful search
→ answer produced
→ corrective feedback
impossible and uncorrected
→ guessing or a blank
→ error may persist
The useful unit is effortful retrieval followed by an accurate answer. A vague feeling of familiarity after flipping the card doesn’t count
2. retention can be measured
In a 2006 experiment, students learned prose through four study periods, three study periods plus one free-recall test, or one study period plus three tests
Five minutes later, repeated study produced the highest recall. One week later, repeated retrieval produced the highest recall
The graph is linked to Roediger and Karpicke’s study. Passive review looked better immediately because the material remained fluent, but retrieval produced slower forgetting
A newer meta-analysis of applied classroom studies found a moderate advantage for distributed over massed practice, $d=0.54$ with a 95% confidence interval from $0.31$ to $0.77$
$$ \text{current fluency} \neq \text{future retrievability} $$
a real decay calculation
A simple descriptive model is:
$$ R(t)=R_0e^{-\lambda t} $$
$R(t)$ is the fraction recalled after $t$ days, $R_0$ is recall near the start of the delay, and $\lambda$ is the fitted decay constant
Using the graph’s 5-minute and 1-week values:
$$ \lambda_{\mathrm{study}} = -\frac{1}{7}\ln\left(\frac{0.40}{0.83}\right) = 0.104\ \mathrm{day^{-1}} $$
$$ \lambda_{\mathrm{retrieval}} = -\frac{1}{7}\ln\left(\frac{0.61}{0.71}\right) = 0.0217\ \mathrm{day^{-1}} $$
Since $t_{1/2}=\ln(2)/\lambda$, the fitted half-lives are 6.7 days for repeated study and 31.9 days for repeated retrieval
These are calculations from real group means, not universal biological constants. Two observations can’t prove an exponential curve between them, and a group recalling prose doesn’t define one person’s memory half-life
The narrow result is still useful: retrieval started lower and lost a much smaller fraction of accessible information over the following week
why spacing matters
Repeating an answer five times in five minutes can leave it in working memory. Waiting introduces some forgetting, changes the study context, and adds interference from other information
The next attempt should be hard enough to require retrieval but early enough to avoid full relearning
$$ \text{too soon} \rightarrow \text{easy but low-information} $$
$$ \text{too late} \rightarrow \text{failure and relearning} $$
A review of 317 experiments found that the best spacing interval generally increased as the delay to the final test increased
Ebbinghaus’s work
Ebbinghaus’s forgetting curve describes rapid loss soon after learning followed by progressively slower forgetting. He measured retention as savings—the percentage of effort avoided when relearning material—and fitted the following equation to his own nonsense-syllable experiments, where b is percentage savings, t is elapsed time in minutes, and the logarithm is base 10. The constants were specific to his data rather than universal values Ebbinghaus’s original analysis
$$b(t)=\frac{100k}{(\log_{10}t)^c+k}$$ $$k=1.84,\quad c=1.25$$
Wozniak’s work
The SM-2 algorithm, developed by Piotr Wozniak, converts recall performance into expanding review intervals. The first successful intervals are 1 and 6 days, after which each interval equals the previous interval multiplied by the item’s easiness factor Wozniak’s original SM-2 description
$$I(1)=1$$ $$I(2)=6$$ $$I(n)=I(n-1)\times EF,\quad n>2$$
After each review, the easiness factor is updated from the response quality q, scored from 0 to 5, with a minimum value of 1.3. Scores below 3 restart the item’s interval sequence
$$EF’=\max\left(1.3,\ EF+0.1-(5-q)\left[0.08+0.02(5-q)\right]\right)$$
3. Anki estimates when a card will be forgotten
Current Anki versions can use the Free Spaced Repetition Scheduler, or FSRS. It tracks three quantities for each card
| Quantity | Meaning |
|---|---|
| difficulty $D$ | how hard the card is to stabilize |
| stability $S$ | days for predicted recall to fall from 100% to 90% |
| retrievability $R$ | probability of recalling the card now |
The Anki manual uses those same definitions
the scheduler’s equation
FSRS-6 models recall as:
$$ R(t,S) = \left( 1+f\frac{t}{S} \right)^{-w_{20}} $$
where:
$$ f=0.9^{-1/w_{20}}-1 $$
With the published default $w_{20}=0.1542$, $f\approx0.9803$. For a card with $S=10$ days, predicted recall is 94.0% after 5 days, 90.0% after 10 days, 84.6% after 20 days, and 74.3% after 60 days
The graph links to the FSRS-6 algorithm specification. It is a scheduler prediction, not a measurement of a neuron or proof that one card has exactly 84.6% recall on day 20
the buttons are data
| Button | What happened before revealing the answer |
|---|---|
| Again | wrong or absent |
| Hard | correct with serious difficulty |
| Good | correct with some hesitation |
| Easy | correct and immediate |
If someone reveals the answer, thinks “I basically knew that,” and presses Good, the log records a successful retrieval that didn’t happen. The scheduler then lengthens the interval using corrupted data
$$ \text{rating} = \text{retrieval before reveal} $$
The answer buttons aren’t grades for effort or punishments for annoying cards. A badly written card should be rewritten or suspended
every card creates future work
The Anki manual estimates average daily load as:
$$ L=\sum_{i=1}^{N}\frac{1}{I_i} $$
$I_i$ is the interval in days for card $i$
If 500 cards have 10-day intervals and 1,500 have 100-day intervals:
$$ L = 500\left(\frac{1}{10}\right) + 1500\left(\frac{1}{100}\right) = 65\ \text{reviews/day} $$
At 8 seconds per mature card, answer time alone is about 8.7 minutes per day. This excludes new cards, failures, editing, and breaks
Adding a card isn’t a one-time note. It is a future review liability
4. Anki fits medicine unusually well
Preclinical medicine contains thousands of discrete associations that must remain accessible after the organ block ends. Anki gives each item its own schedule and turns “review cardiology” into a finite queue
In a survey of 560 students from 102 US medical schools, 68.3% reported using Anki, making it one of the most common third-party resources in the sample
The multi-school survey helps explain the reverence but doesn’t prove effectiveness. Shared decks reduce setup time, online communities correct mistakes, and daily streaks make adherence visible
what USMLE studies found
| Study | Result | Limit |
|---|---|---|
| 165 students at one institution | daily use correlated with higher Step 1 scores, $p=.039$ | no Step 2 association and no randomization |
| earlier survey of 72 students | about 1 additional Step 1 point per 1,700 completed cards after adjustment | card volume may track student consistency |
| survey-linked Step 1 study | more consistent use correlated with higher scores and perceived retention | self-selection remains |
The 165-student study also found higher reported sleep quality among daily users, while other wellness outcomes didn’t differ
The separate Step 1 study recommended Anki alongside question banks and called for research on later clinical performance
These studies used historical numeric Step 1 scores before the exam became pass/fail. They show association, not that Anki caused a particular score
retrieval practice has causal experimental support
+ spacing has causal experimental support
+ Anki implements both at scale
+ Anki use correlates with some USMLE outcomes
- Anki itself hasn't been shown to cause a specific Step score
5. cards don’t replace exam performance
Retrieval practice works outside medicine. A meta-analysis of 19 publications in psychology classes found a benefit of about $d=0.56$ across 72 effect sizes
The classroom meta-analysis is closer to real exams than a one-session word-list study
In undegrad programming, students exposed to more frequent testing later scored 9.1 to 13.5 percentage points higher on code-writing questions than students in an earlier infrequent-testing semester. That study measured an applied skill, but the separate cohorts prevent us finding any causation
6. treadmill Anki is a dual task
Two claims often get merged
acute exercise can change arousal, attention, and memory
doing Anki during exercise improves Anki retentionThe first has experimental support under some conditions. The second hasn’t been directly established well enough to treat treadmill Anki as a special memory intervention
In one study, 88 college students rested or completed a 15-minute moderate treadmill walk before encoding, during encoding, or during early consolidation. The memory effects depended on exercise timing
Another experiment found better immediate and 24-hour memory after moderate exercise in 23 young adults, but 20 versus 40 minutes didn’t change the overall benefit. That experiment tested exercise around the memory task rather than a long card queue during walking
Exercise can change catecholamines, lactate, arousal, blood flow, and circulating factors associated with plasticity. A rise in peripheral BDNF doesn’t prove that a specific flashcard formed a stronger cortical trace
Simultaneous walking is also a dual task. As speed, incline, balance demand, card complexity, or fatigue rises, gait and cognition compete more strongly for attention
| Treadmill condition | Reasonable use |
|---|---|
| slow, stable walking | mature cards with short verbal answers |
| moderate effort | simple review only if accuracy remains stable |
| high incline or running | no screen-based review |
| new mechanisms or calculations | seated study with scratch space |
| repeated misses | stop walking or rewrite the card |
7. the deck should be tested against reality
write extremely scorable cards and be objective throughout a session
A scheduler can time a bad card perfectly, but it still is just as useless as running a random shuffle
Useful cards have one clear cue, one answer that can be judged, enough context to identify the concept, and a response small enough to retrieve without turning into an essay
| Weak card | Better target |
|---|---|
| “Tell me everything about heart failure” | one mechanism, criterion, or comparison |
| isolated drug name | drug plus indication, contraindication, or adverse-effect cue |
| copied paragraph | one direct question with a short answer |
| long cloze chain | separate cues that require the missing relation |
| trick wording | direct wording with variation practiced elsewhere |
atomic doesn’t mean context-free. “What does metoprolol do” is short but underspecified. “Which receptor does metoprolol preferentially block at ordinary doses” is scorable, ideally 1-4 to assign it to each aptitude level (good, med, hard, repeat)
protect sleep
Sleep supports consolidation and next-day attention. A tired learner may take longer, fail more mature cards, and generate extra relearning work that appears over the next several days
$$ \text{less sleep} \rightarrow \text{worse encoding and retrieval} \rightarrow \text{more lapses} \rightarrow \text{larger future queue} $$
measure transfer
Daily percentage correct mixes new, young, mature, easy, and hard cards. Better measurements include:
mature-card true retention
retention by interval
median response time
lapse rate
leech rate
reviews/minutes per day (old, known)
accuracy on unseen practice questionsIf card retention rises while unseen-question performance stays flat, the deck may be training its own cues or preserving facts that aren’t limiting exam performance
For treadmill reviews, compare matched decks:
$$ \Delta R = R_{\mathrm{walk}} – R_{\mathrm{seated}} $$
$$ \Delta T = T_{\mathrm{walk}} – T_{\mathrm{seated}} $$
$R$ is later retention and $T$ is median response time. More cards completed isn’t a benefit if retention falls or response time erases the time saved
8. the main conclusion
Active recall trains the operation an exam and real-world decision often require: recovering usable information from an incomplete cue
Spacing adds controlled forgetting so retrieval remains effortful without turning every review into full relearning. Anki is revered in pre-med and medicine because it can run that process across thousands of individually scheduled facts, not because its interface has a unique effect on the brain
The best causal evidence supports retrieval practice and spacing in general. Anki’s USMLE evidence is encouraging but observational, and direct proof for LSAT or bar passage is too weak to claim. Slow treadmill review is reasonable for simple mature cards if later retention and response time remain comparable with seated review, but exercise research doesn’t prove a special combined effect
TLDR. My personal learning: active recall means producing an answer before seeing it, while spacing decides when to demand that answer again. Retrieval can reinstate hippocampal–cortical memory patterns and creates another chance to strengthen or correct the cue–answer relationship. In a classic prose experiment, repeated study produced 83% immediate recall but fell to 40% after one week, while repeated retrieval started at 71% and retained 61%. A two-point exponential fit gives descriptive half-lives of 6.7 versus 31.9 days, but those aren’t universal constants. Anki’s FSRS scheduler estimates difficulty, stability, and recall probability for each card, then uses honest review outcomes to update the next interval. Anki fits medicine because the fact volume is large and cumulative, and its use correlates with some historical USMLE Step 1 scores without proving causation. Retrieval and spacing improve learning outside medicine too, but strong direct evidence for Anki on the LSAT or bar isn’t available. Slow treadmill review is reasonable for simple mature cards if later retention and answer time stay comparable with seated review. The best system uses cards for durable components, practice questions for transfer, sleep for consolidation, and retention data instead of streaks or cards completed
Comments are closed for this post.