Engineers learn for a living. A new framework every year, a new codebase every job, a cloud certification, a database you suddenly own, a paper the team wants to discuss on Friday. Most of us learn it the way we learned in school: read the docs, reread the tricky parts, highlight, maybe take notes, and hope it sticks.

It mostly does not. A week later the details are gone, and what remains is a feeling of familiarity that collapses the moment you need to write the code without the docs open.

Cognitive psychology has studied this for over a century, and two findings are among the most robust in the field: testing yourself beats rereading, and spreading practice out beats cramming. This post covers what those mean, the experiments behind them with their real numbers, where the evidence is weaker, and how to turn them into a routine that fits an engineering week.

1. The vocabulary

Term Meaning Why it matters
Retrieval practice / testing effect Pulling information out of memory (answering a question, writing code from scratch) instead of putting it back in (rereading) The act of recall itself strengthens memory. It is a learning tool, not only a measurement.
Spacing / distributed practice Spreading study of the same material over separate sessions Same total time, much better long-term retention than one massed session.
Massed practice / cramming All study in one block Feels efficient, gives good short-term performance, fades fast.
Interleaving Mixing different problem types or topics within a session Trains you to choose the right method, not just execute the one you were told to use.
Desirable difficulty A condition that makes learning feel harder and slower but improves long-term retention and transfer Explains why the best techniques feel worse while you use them.
Fluency illusion Mistaking ease of processing (it reads smoothly, it looks familiar) for knowing it The main reason people keep rereading.
Retention interval Time between study and when you need the knowledge Determines how far apart reviews should be.
Spaced repetition system (SRS) Software that schedules reviews at growing intervals based on your answers Automates spacing and retrieval together. Anki is the best-known example.

2. Retrieval practice: the evidence

Testing beats restudying, but only later

Roediger and Karpicke (2006) had students read short prose passages and then either reread them or take a free-recall test, with no feedback. The results flip depending on when you measure:

Final test delay Restudy Tested once Takeaway
5 minutes 81% 75% Rereading wins in the short term
2 days 54% 68% Testing wins
1 week 42% 56% Testing wins by more

In a second experiment, with four sessions, the pattern sharpened. After one week, students who studied once and recalled three times (STTT) remembered 61%, versus 40% for those who studied four times (SSSS). Five minutes after learning, the order was reversed. And the SSSS group was the most confident it would remember the passage.

That is the whole problem in one result: the strategy that feels best and performs best immediately is the one that loses a week later.

Once you know it, keep retrieving it

Karpicke and Roediger (2008) taught students 40 Swahili-English word pairs. Once a student recalled a pair correctly, it was either kept in future tests or dropped from them, and either kept in future study or dropped.

  • Keeping items in testing led to about 80% recall a week later.
  • Dropping items from testing led to 33-36%, even when students kept studying them.
  • Continuing to study items already recalled had no measurable benefit.
  • Students in every condition predicted they would remember about 50%; predictions did not track the real outcome.

Engineering translation: "I got it right once" is not done. The common habit of dropping a flashcard or skipping a topic after one success throws away most of the benefit.

Retrieval beats "deeper" study methods too

Karpicke and Blunt (2011) compared retrieval practice with concept mapping, a method usually considered deep and active. A week later, on questions that included inference, retrieval practice scored about 50% higher than concept mapping (0.67 vs. 0.45 proportion correct in the first experiment). In the second experiment, 84% of students did better with retrieval practice, while 75% predicted concept mapping would work as well or better.

How big is the effect overall?

Rowland's (2014) meta-analysis pooled 159 effect sizes from 61 studies comparing testing with restudying:

Condition Effect size (Hedges' g)
Overall 0.50
With feedback after the test 0.73
Without feedback 0.39
Retention interval of 1 day or more 0.69
Retention interval under 1 day 0.41

A g of 0.5 is a medium effect by common conventions and large for an intervention that costs nothing. Two practical points come straight from the table: check your answers (feedback nearly doubles the effect), and the benefit grows with time, which is exactly when you need the knowledge.

Recall attempts help even when they fail. Kornell, Hays and Bjork (2009) found that trying to answer a question before seeing the answer improved later memory compared with simply studying the question and answer together, even when the attempt was wrong. Guessing first, then checking, is not wasted time.

3. Spacing: the evidence

The spacing effect is huge and old

The forgetting curve goes back to Hermann Ebbinghaus (1885), who memorized nonsense syllables and measured how fast he forgot them. Murre and Dros (2015) replicated the experiment with one subject over 70 hours of learning and relearning, and found a curve very similar to the original, with an odd upward bump around 24 hours that may reflect sleep.

Cepeda et al. (2006) reviewed 839 assessments of distributed practice from 317 experiments in 184 articles. Spaced study beat massed study reliably, and the optimal gap between sessions grew as the retention interval grew.

How far apart should reviews be?

Cepeda et al. (2008) tested more than 1,350 people with gaps of up to 3.5 months between study sessions and final tests up to a year later. The optimal gap, as a fraction of the time until the test, declined from about 20-40% for a 1-week test to about 5-10% for a 1-year test.

You need it in Rough first review gap (from that ratio)
1 week (a demo, an interview) 1-3 days
1 month (a certification exam) about a week
1 year (a language you will use at work) 3-5 weeks for a second session, then keep expanding

Two caveats: these are ratios from verbal learning studies, not laws, and the curve is flat near its peak. Being somewhat too late costs less than being too early, and almost any spacing beats none.

Learners prefer cramming even when it fails them

Kornell (2009) had students learn vocabulary with flashcards in one big spaced stack or several small massed stacks. Spacing was better for 90% of participants, yet after the first session 72% believed massing had worked better. Kornell and Bjork (2008) found a similar illusion for interleaving. Your sense of what is working is not a reliable guide.

4. Interleaving and desirable difficulties

Robert Bjork's idea of desirable difficulties (Bjork, 1994; Bjork and Bjork, 2011) ties these findings together: conditions that slow down apparent progress, such as retrieval, spacing and variation, often improve long-term retention and transfer. The difficulty has to be one you can overcome; a test on material you never studied is just hard.

Interleaving is the less obvious member of the family:

  • Rohrer and Taylor (2007): students practiced volume problems for different solids either blocked by type or shuffled. During practice the blocked group scored higher (89% vs. 60%). On a test a week later, the shuffled group scored 63% vs. 20%.
  • Kornell and Bjork (2008): people learning painters' styles classified new paintings better after seeing artists interleaved rather than one artist at a time, yet most still rated blocking as more effective.

Why it works for engineering material: real problems do not come labeled. Blocked practice teaches "this is a chapter 4 problem, use the chapter 4 method". Interleaving teaches you to recognize which tool fits, which is most of the skill.

5. What the evidence says about popular study methods

Dunlosky et al. (2013) reviewed ten common learning techniques for Psychological Science in the Public Interest, rating them on how general and robust the evidence was:

Utility rating Techniques
High Practice testing, distributed practice
Moderate Elaborative interrogation (asking "why is this true?"), self-explanation, interleaved practice
Low Summarization, highlighting, keyword mnemonic, imagery for text, rereading

Low utility does not mean useless; it means weak or narrow evidence relative to time spent. Rereading and highlighting are what most people do most.

6. Where the evidence is weaker

Being honest about limits makes the strong results more credible:

  • Most studies use verbal material. Word pairs, prose passages and facts dominate. The effects extend to concepts and inference (Karpicke and Blunt) and to math problems (Rohrer and Taylor), but there is less direct evidence for complex procedural skills like debugging or system design.
  • Classroom results are more mixed than lab results. Bego et al. (2024) tested spaced retrieval practice across nine introductory STEM courses and found a significant benefit in only two; the combined effect depended on which courses were included. Real courses add noise: compliance, varied exams, students who already study well.
  • Effect sizes vary with feedback and delay. Rowland's moderators show the testing effect is smaller without feedback and at short delays.
  • Spacing ratios are approximate. The 20-40% and 5-10% figures come from specific materials and tests. Treat them as starting points.

The direction of the effect is about as solid as psychology gets. The size of the effect for your specific skill is an empirical question.

7. A playbook for engineers

Turn reading into retrieval

Instead of Do
Rereading the docs page Close it, write down the API from memory, then check
Following a tutorial step by step Do it once, then rebuild it from a blank file the next day
Highlighting a design doc Write three questions it answers, answer them a day later without looking
Watching a conference talk Pause at the end and write the five main claims before checking the slides
Reading a codebase top to bottom Draw the request path from memory, then trace it in the code to correct it

A weekly routine for learning a new language or framework

  • Day 1: learn a small slice (one concept, one API). Close the docs and write a minimal example from memory. Check and fix.
  • Day 2-3: rebuild yesterday's example from scratch, then add a new slice.
  • End of week: a mixed session. Small exercises that need several concepts, in random order, without notes.
  • Week 2 and 4: revisit the hardest items. Anything you failed goes back in the queue.
  • Always: check against the source after each attempt. Feedback is where much of the effect comes from.

Spaced repetition, used well

Spaced repetition software automates both principles. Anki 23.10 (October 2023) integrated the FSRS scheduler, which fits review intervals to your own recall history instead of using the older fixed SM-2 rules.

Good cards for engineers:

  • One fact per card. "What does git rebase --onto A B C do?" beats "Explain git rebase".
  • Cloze code. A real snippet with one argument blanked out: SELECT ... FROM orders ____ JOIN customers USING (customer_id).
  • Why cards. "Why is a B-tree index useless for LIKE '%abc'?" Retrieval of reasons transfers better than retrieval of trivia.
  • Commands and flags you actually use. Not the whole man page.
  • Do not drop a card after one correct answer. Let the scheduler decide; Karpicke and Roediger is the reason.

Things SRS does not replace: building something real, reading code, and interleaved practice on full problems. Use cards for the vocabulary of a domain and projects for the skill.

Use it for the team

  • Onboarding: replace "read these 12 pages" with a short set of questions new joiners answer from memory at the end of day 1, day 3 and week 2.
  • Incident reviews: a week later, ask the team to recall the root cause and the fix before rereading the write-up.
  • Reading groups: open with five minutes of "what do you remember from last session?" before the new paper.
  • Interview prep and certifications: mixed practice sets, spaced over weeks, with answers checked every time.

Why this matters

  • Retrieval practice is one of the strongest findings in learning research. A medium effect overall (g of about 0.5), larger with feedback and longer delays.
  • Spacing multiplies it. The same hours spread over days and weeks are worth much more than one long session.
  • Your intuition is biased toward the wrong methods. Rereading, cramming and blocking feel better while you do them and lose a week later.
  • Interleaving trains the real skill. Choosing the right approach is most of engineering, and blocked practice never exercises it.
  • The evidence has limits. Lab results are stronger than classroom results, and complex skills are less studied, so measure on yourself.

Final Thoughts

The uncomfortable part of this research is that the effective methods feel worse. Closing the docs and failing to recall an API feels like not learning. Rereading feels like progress. The experiments say the opposite, over and over, and they also say you will not believe it until you try.

So try it for a month on one thing you actually need: a framework, a codebase, an exam. Recall before you reread, space your sessions, mix your problems, check every answer. Then test yourself at the end, which, conveniently, is also practice.

Sources