Desirable Difficulties: Why Making Learning Harder Helps You Remember More

Introduction

In 1978, a psychologist at the University of Rochester handed college students a simple task. Half of them read word pairs like "rapid - fast." The other half saw "rapid - f___" and had to fill in the blank themselves. The second group struggled more. They were slower. They made errors. And on a surprise memory test one week later, they remembered far more [1]. The finding baffled everyone, including the researcher who ran the experiment. How could making something harder make it stick better? Sixteen years later, cognitive psychologist Robert A. Bjork at UCLA gave this paradox a name. He called it "desirable difficulties" [2]. The term captured a pattern that had been hiding in plain sight across a century of memory research. Spacing your study sessions apart instead of cramming. Mixing different topics instead of practicing one at a time. Testing yourself instead of rereading. All of these feel harder. All of them produce worse performance during practice. And all of them lead to dramatically better learning in the long run. This is the story of how that paradox was discovered, why the brain works this way, and what it means for anyone trying to learn anything that matters.

The Psychologist Who Said Forgetting Is Your Friend

The story of desirable difficulties starts with forgetting. And forgetting starts with Hermann Ebbinghaus.

In 1885, working alone in a small apartment in Berlin, Ebbinghaus did something no one had attempted before. He turned memory into a science. He memorized lists of nonsense syllables, three-letter combinations like DAX, BUP, and ZOL, chosen specifically because they carried no meaning. Then he tested himself at precise intervals. Hours later. Days later. Weeks later. And he plotted the results [3].

The result was the forgetting curve. Memory dropped fast at first, then leveled off. Within twenty minutes, roughly 40% was gone. Within a day, about 67%. Within a month, nearly 80%. But Ebbinghaus noticed something else. When he relearned forgotten material, it came back faster each time. And when he spaced his study sessions apart instead of cramming them together, retention improved. He called this the spacing effect.

For over a hundred years after Ebbinghaus, researchers kept rediscovering pieces of the same puzzle. In 1917, Arthur Gates at Columbia University found that students who spent most of their study time reciting material from memory outperformed those who spent the same time rereading it. In 1966, William Battig observed that mixing different tasks during motor learning slowed acquisition but improved transfer. In 1979, John Shea and Robyn Morgan showed that randomly ordering three motor tasks during practice produced worse performance during training but better retention ten days later [4].

Each finding seemed isolated. Spacing. Testing. Mixing. Generating. Different researchers, different labs, different decades. Nobody had connected the dots.

That changed in 1994 when Robert Bjork published a chapter titled "Memory and Metamemory Considerations in the Training of Human Beings" in a book on metacognition edited by Janet Metcalfe and Arthur Shimamura. In thirty pages, Bjork wove together a century of scattered findings into a single framework. His central claim was deceptively simple. Certain manipulations that slow down learning in the short term actually speed it up in the long term. He called these manipulations desirable difficulties [2].

The word "desirable" was doing real work. Not all difficulties help. Confusing instructions, noisy classrooms, and impossible tasks are just difficulties. They hurt learning. A difficulty is desirable only when it forces the learner to engage in deeper cognitive processing, and only when the learner has enough background knowledge to respond to it successfully.

But why would the brain work this way? Why would struggle produce better memory than ease?

Two Strengths, One Memory

The answer came from Bjork himself, working with his wife and collaborator Elizabeth Ligon Bjork. Together, they built the theoretical engine that powers the entire desirable-difficulties framework. They called it the New Theory of Disuse [5].

The theory starts with a distinction that seems technical but changes everything. Every memory, they argued, has two independent measures. Storage strength is how well a memory is woven into the web of everything else you know. Retrieval strength is how easily you can access it right now.

Think of it like a book in a library. Storage strength is how many cross-references that book has to other books. Retrieval strength is whether you can find it on the shelf this instant.

Here is the counterintuitive part. Storage strength only goes up. Once something is deeply encoded, it stays encoded. You never truly erase a memory. What you lose is retrieval strength. That fades with time, interference, and disuse. And here is the key: when retrieval strength is low, successfully retrieving the memory produces the largest gains in storage strength [6].

Read that again. When something is hard to recall, the act of successfully recalling it strengthens the underlying memory trace more than when it is easy to recall. The struggle is not a side effect. The struggle is the mechanism.

This single principle explains why all desirable difficulties work. Spacing creates a gap that lets retrieval strength decay, so the next retrieval attempt is harder but produces a bigger storage-strength boost. Interleaving forces discrimination between similar items, requiring more effortful processing. Testing forces active reconstruction rather than passive recognition. Generation forces the learner to build the answer rather than simply receive it.

Soderstrom and Bjork confirmed this framework in a sweeping integrative review in 2015 [7]. They catalogued dozens of studies showing that learning and performance can be not merely uncorrelated but inversely related. The conditions that make you perform best during practice are often the conditions that produce the least durable learning.

Or as Elizabeth Bjork puts it: "Forgetting is the friend of learning."

What does this mean for anyone studying for an exam or learning a new skill? It means that the feeling of fluency during study is a trap. When material feels easy to process, retrieval strength is high but storage strength is not necessarily increasing. When material feels difficult, that discomfort is the signal that real learning is happening.

Five Ways to Make Learning Harder (and Better)

The desirable-difficulties framework identifies five core techniques supported by decades of experimental evidence. Each one makes practice harder. Each one makes learning last.

The first and most studied is spacing. The idea is simple: spread your study sessions apart instead of massing them together. The evidence is overwhelming. In 2006, Nicholas Cepeda and colleagues published the definitive meta-analysis. They analyzed 184 articles containing 317 experiments. The conclusion: spaced practice virtually always beats massed practice for any retention interval beyond a few minutes [8]. A follow-up study by the same group in 2008 found that the optimal spacing gap is roughly 10 to 20 percent of the target retention interval [9]. Studying for an exam in two months? Space your sessions about five to ten days apart.

The second is interleaving. Instead of practicing one type of problem until you master it before moving to the next (blocked practice), you mix different types together. This feels chaotic. Performance during practice drops. But long-term retention and transfer improve. Kornell and Bjork showed in 2008 that participants who studied paintings by twelve artists in interleaved order roughly doubled their accuracy at identifying new paintings compared to those who studied in blocked order [10]. Doug Rohrer's team has confirmed this in real classrooms. In a preregistered randomized trial with 7th-grade mathematics students, interleaved practice produced large and lasting improvements on delayed tests [11].

Brunmair and Richter published the most thorough meta-analysis of interleaving in 2019, covering 59 studies and 238 effect sizes. The overall effect was g = 0.42 [12]. But the moderation pattern matters. Effects are strongest for visual category learning (g around 0.67) and mathematics (g around 0.34). They reverse for unrelated word lists (g around -0.39). Interleaving works when categories share enough features that comparison reveals meaningful differences. When items are unrelated, mixing them just creates noise.

The third is retrieval practice. This is testing yourself, not to assess what you know, but to strengthen what you know. Roediger and Karpicke ran the landmark experiment in 2006. Students who studied a passage once and then took three free-recall tests remembered far more after one week than students who restudied the passage four times [13]. Karpicke and Blunt extended this in 2011 in a study published in Science, showing that retrieval practice outperformed even sophisticated elaborative studying with concept mapping [14]. A 2017 meta-analysis by Adesope and colleagues, aggregating 272 effect sizes from 188 experiments, found a mean effect of g around 0.51 compared to restudy [15].

The fourth is the generation effect. When you generate an answer rather than reading it, memory improves. Slamecka and Graf demonstrated this systematically in 1978 [1]. The effect holds across formats: word stems, anagrams, definitions, and problem solving. A meta-analysis found a mean effect of d around 0.40 across 86 studies. Even unsuccessful generation attempts followed by feedback enhance later learning, a finding Kornell, Hays, and Bjork confirmed in 2009.

The fifth is varying the conditions of practice. In 1978, Smith, Glenberg, and Bjork showed that studying the same material in two different rooms produced about 50 percent higher free recall than studying it twice in the same room [16]. The explanation is that variable contexts produce memory representations tied to multiple cues rather than a single context. When the test environment is unpredictable (the realistic case for most real-world performance), variable encoding gives you more retrieval routes.

Inside the Brain: Why Struggle Strengthens Synapses

The behavioral evidence for desirable difficulties is overwhelming. But what is happening at the neural level? Why does the brain encode effortful memories more deeply than effortless ones?

The answer starts at the synapse. The cellular mechanism of long-term memory is long-term potentiation, or LTP. First described by Bliss and Lomo in 1973 in the rabbit hippocampus, LTP is a persistent strengthening of synaptic transmission following high-frequency stimulation [17]. In 2006, Whitlock and colleagues published a paper in Science showing that behavioral learning induces the same molecular changes as artificially induced LTP in the hippocampus [18].

Here is the connection to desirable difficulties. LTP requires above-threshold postsynaptic activation. Casual, low-effort exposure does not reliably reach that threshold. Effortful retrieval, which requires strong and sustained neural activation, does. The late phase of LTP (L-LTP) requires protein synthesis and gene transcription that takes hours to consolidate. Spaced repetition fits this timeline perfectly: each retrieval event triggers a fresh consolidation cascade, whereas massed repetition piles new learning onto a still-labile trace.

The hippocampus, that seahorse-shaped structure deep in the temporal lobe, plays a central role. The standard model of memory consolidation holds that the hippocampus rapidly binds new associations and then gradually transfers them to slower-learning neocortical networks during offline periods, especially sleep. During slow-wave sleep, hippocampal sharp-wave ripples replay recent encoding sequences [19]. This replay is critical for transforming temporary hippocampal traces into stable neocortical representations. Diekelmann and Born's influential 2010 review in Nature Reviews Neuroscience confirmed that sleep between study sessions is mechanistically central to spaced learning [20].

Spacing gaps allow the hippocampus to replay and the cortex to integrate before the next study trial arrives. Without gaps, each new encoding overwrites the previous one before consolidation completes.

The prefrontal cortex adds another piece. Functional neuroimaging consistently shows that effortful retrieval recruits anterior prefrontal cortex more strongly than passive restudy. Lepage and colleagues at McGill showed in 2000 that the right anterior prefrontal cortex, particularly Brodmann area 10, activates specifically during retrieval attempts [21]. Liu and colleagues demonstrated in 2020, in a study published in eLife, that retrieval practice produces more differentiated and stronger representations in medial prefrontal cortex compared to restudy [22].

Then there is dopamine. Dopaminergic neurons in the ventral tegmental area encode reward prediction error, the difference between expected and actual outcomes. Tanaka and colleagues showed in 2019 that this prediction-error signal is amplified after effortful action [23]. The brain treats hard-won retrievals as more valuable and tags their memory traces for stronger consolidation. This explains why immediate feedback after effortful retrieval produces bigger gains than feedback after fluent retrieval.

Finally, there is reconsolidation. In 2000, Nader, Schafe, and LeDoux demonstrated that retrieving an established memory transiently destabilizes it, requiring new protein synthesis to re-stabilize [24]. This reconsolidation window means each retrieval is not a passive readout. It is an opportunity for the trace to be re-encoded with new context, additional cues, and stronger connections. Fukushima and colleagues confirmed in 2014 that retrieval can actively enhance memory through this reconsolidation pathway [25].

What does this mean? Every time you struggle to recall something and succeed, your brain is doing at least four things simultaneously. It is strengthening synaptic connections through LTP. It is engaging prefrontal monitoring circuits that deepen encoding. It is triggering dopamine signals that tag the memory as important. And it is opening a reconsolidation window that allows the trace to be modified and strengthened. None of these processes fire with the same intensity during passive rereading.

The Illusion That Keeps Learners Trapped

If desirable difficulties work so well, why does almost nobody use them voluntarily?

The answer is one of the most important findings in the entire field. Learners systematically prefer strategies that feel easy and reject strategies that work.

Kornell and Bjork demonstrated this in 2007. Even after participants had just outperformed themselves under interleaved conditions, the majority insisted that blocked practice had been more effective [26]. In a separate study in 2008, 78 percent of participants rated blocking as superior to interleaving for learning, despite performing better with interleaving [10].

The culprit is what Benjamin, Bjork, and Schwartz called the "fluency illusion" in a landmark 1998 paper. Judgments of learning track current retrieval fluency, how easily material comes to mind right now. But fluency at encoding is a poor predictor of long-term retention. In many cases, it is inversely related. Material that feels smooth and easy during study produces inflated confidence but weak memory traces [27].

Kirk-Johnson, Galla, and Fraundorf refined this in 2019 with the "misinterpreted-effort hypothesis." They showed that students who perceived a strategy as more effortful rated it as less effective, even though choosing the more effortful strategy was associated with better long-term retention. The effort itself was being misread as failure.

Bjork, Dunlosky, and Kornell brought the entire picture together in a 2013 review in the Annual Review of Psychology [28]. Students overwhelmingly rely on rereading and highlighting, which produce fluency illusions. They underuse retrieval practice and spacing, which feel harder but work. The Dunning-Kruger pattern makes things worse: the least-skilled learners are the most overconfident and the least likely to adopt effective strategies.

This is not a minor problem. It means that left to their own devices, most learners will naturally drift toward the exact strategies that feel best and work worst. The implications for education are profound. Teaching students about desirable difficulties may be as important as teaching the content itself.

When Difficulty Stops Being Desirable

The word "desirable" is conditional. Not every difficulty helps. The boundaries matter as much as the principles.

The most important boundary comes from Cognitive Load Theory, developed by John Sweller beginning in 1988. Working memory has strict capacity limits, roughly four items that can be processed simultaneously. When instructional material is already complex, with many interacting elements that must be understood together, adding further difficulty through interleaving or generation can push total cognitive load past the threshold. The result is not deeper learning but cognitive collapse [29].

Chen, Kalyuga, and Sweller demonstrated this directly in a series of studies between 2015 and 2018. When materials had high element interactivity, the testing effect and generation effect disappeared or reversed. Learners who received worked examples outperformed those forced to generate solutions. The difficulty had become undesirable because there was not enough working-memory capacity left for the effortful processing to succeed.

A related boundary is the expertise-reversal effect, documented by Kalyuga, Ayres, Chandler, and Sweller in 2003 [30]. Instructional supports that help novices, like worked examples and step-by-step guidance, become redundant and even harmful for experts. The same logic runs in the other direction: desirable difficulties that benefit experienced learners can overwhelm beginners. A novice medical student facing interleaved differential diagnosis cases needs to first build basic schemata through structured instruction. An experienced clinician benefits from the discrimination practice that interleaving provides.

Spacing has its own boundary. Cepeda and colleagues showed that the optimal inter-study gap is bounded above as well as below. Gaps far exceeding 20 percent of the target retention interval degrade performance because the memory trace becomes so weak that retrieval fails rather than succeeds. Failed retrieval without subsequent feedback does not strengthen the trace.

Interleaving has limits too. Brunmair and Richter's meta-analysis revealed that interleaving actually hurts for unrelated word lists, with a negative effect of g around -0.39 [12]. The reason: interleaving works by enabling discriminative contrast between similar categories. When categories share no relevant features, juxtaposition produces only confusion.

What does this mean practically? The art of applying desirable difficulties is calibration. Start with clear instruction and worked examples for novices. As knowledge builds and element interactivity drops, gradually introduce spacing, testing, and interleaving. Monitor whether retrieval attempts succeed. If learners are consistently failing, the difficulty has crossed from desirable to destructive.

From Lab Benches to Real Classrooms

Desirable difficulties have been tested far beyond the psychology laboratory. The evidence from real-world settings confirms the core findings while adding important nuance.

In medical education, Doug Larsen, Andrew Butler, and Henry Roediger ran a randomized controlled trial in 2009. Medical residents studied neurology content, status epilepticus and myasthenia gravis, either through repeated testing or repeated study. At six months, the testing group retained significantly more [31]. A 2024 systematic review confirmed these benefits across medicine, physiotherapy, and clinical psychology, while noting practical barriers to implementation [32].

In mathematics, Rohrer's classroom research program has produced some of the strongest real-world evidence. A 2020 cluster-randomized trial with 54 seventh-grade classes showed that interleaved practice assignments produced substantial and lasting improvements on unannounced cumulative tests, even months after the study period ended [11]. A 2025 meta-analytic review confirmed robust benefits of spacing and retrieval practice for mathematics learning across multiple settings [33].

In language learning, spaced practice has been confirmed in a meta-analysis by Kim and Webb covering second-language vocabulary acquisition [34]. Talker variability, a form of varying practice conditions, has been shown to improve phonological learning even for artificial grammars [35]. Suzuki, Nakata, and DeKeyser proposed desirable difficulties as a theoretical foundation for second-language practice in a 2019 review.

In motor learning and sports, the contextual-interference effect originally demonstrated by Shea and Morgan has been extended to surgical skills, bimanual coordination tasks [36], and athletic performance. Wulf and Shea noted an important caveat in a 2002 review: for highly complex tasks, blocked practice may be needed initially to prevent coordinative overload before interleaving can provide benefits.

The Bjork Learning and Forgetting Lab at UCLA continues to be the primary institutional home for this line of research [37]. Robert Bjork's involvement chairing the National Research Council Committee on Techniques for the Enhancement of Human Performance from 1988 to 1994 reflects the long-standing interest of military and aviation training communities in these principles.

The Debate That Will Not Settle

Despite the strength of the evidence, desirable difficulties face legitimate criticism.

The most persistent is the circularity objection. Critics, including Jacoby, Wahlheim, and Coane in a 2010 paper, argued that the theory risks becoming unfalsifiable. When a difficulty helps, it was desirable. When it hurts, it was undesirable. The definition appears to depend on the outcome [38].

The reply is that the framework does specify in advance when difficulties will be desirable, based on learner expertise, element interactivity, material similarity, and retention interval. These predictions have been confirmed by Brunmair and Richter's moderation analysis, Kalyuga's expertise-reversal program, and the Cepeda spacing function. The framework is not circular. It is conditional.

A second concern is effect-size heterogeneity. While the broad effects replicate, individual studies vary widely. Some classroom interleaving trials show null results. A recent large-scale randomized trial found significant effects on within-year skills tests but null effects on year-end cumulative assessments, raising questions about whether lab-scale findings always translate under realistic curriculum constraints.

A third challenge is implementation. Bjork and Bjork opened a 2020 special issue of the Journal of Applied Research in Memory and Cognition by noting the persistent gap between scientific evidence and educational practice [39]. Textbooks are organized in blocked fashion. Grading systems reward immediate performance. Students and parents expect smooth, confidence-building practice. All of this works against desirable difficulties. Biwer and colleagues described interventions that explicitly teach students about the framework, with generally positive results.

A subtler theoretical question concerns the storage-strength and retrieval-strength distinction itself. Some computational models replace this with continuous decay parameters and rate-of-forgetting functions. These models are largely compatible with the qualitative predictions of the New Theory of Disuse but propose different mechanistic accounts.

None of these critiques overturns the core findings. The spacing effect has been replicated hundreds of times across more than a century. The testing effect is one of the most robust phenomena in experimental psychology. Interleaving has clear moderators that predict when it works and when it does not. What the critiques show is that desirable difficulties are not a magic formula. They are a family of conditional principles that require thoughtful calibration.

A Timeline of Discovery

Eight Principles That Stick

Distilling decades of research into practical guidance yields a clear set of principles.

Space your study sessions. Use intervals of roughly 10 to 20 percent of your target retention interval. For an exam in two months, study sessions five to ten days apart are near optimal. For material you want to retain for a year, gaps of weeks to a month become appropriate.

Test yourself instead of rereading. Free recall, short-answer questions, and elaborative cued recall produce the largest effects. Always check your answers afterward. Feedback amplifies the benefit, especially after difficult retrieval attempts.

Interleave related but confusable material. Mix problem types whose appropriate strategies are not obvious from surface features. Do not interleave material that is completely unrelated. The power of interleaving comes from forced discrimination, not random mixing.

Generate before you receive. Try to answer a question, predict an outcome, or solve a problem before reading the explanation. Even wrong answers followed by feedback enhance later learning. This is sometimes called the pretesting effect.

Vary the conditions of practice. Study in different locations, at different times of day, using different formats. Rephrase ideas in different words. Tackle problems in different representations. Variability builds context-independent memory representations.

Calibrate difficulty to expertise. For complete beginners, prioritize clear explanations and worked examples. As knowledge builds and working memory frees up, gradually introduce spacing, testing, and interleaving. If retrieval attempts are consistently failing, reduce the difficulty.

Trust outcomes over feelings. The subjective sense of ease during study is misleading. Use delayed, low-stakes self-tests to assess actual learning. If practice feels smooth, that is a warning sign, not a reassurance.

Respect the role of sleep. Spacing benefits depend in part on sleep-mediated consolidation. Late-night cramming defeats both the spacing effect and the hippocampal replay it relies on. Study, sleep, retrieve, sleep, retrieve. That sequence works with the brain's architecture rather than against it.

Conclusion

The story of desirable difficulties is, at its heart, a story about trust. Trusting that struggle is not a sign of failure but a signal of growth. Trusting that forgetting is not the enemy of memory but its precondition. Trusting that the brain knows how to learn, even when consciousness insists otherwise.

Robert Bjork began this story in 1994 with a simple observation: the conditions that produce the best performance during practice are not the conditions that produce the best learning. Three decades and hundreds of experiments later, that observation has become one of the most robust findings in cognitive science. It has been confirmed in laboratories and classrooms, with children and adults, in medicine and mathematics and music and military training.

The neuroscience has caught up to the psychology. Effortful retrieval strengthens synapses through long-term potentiation. Sleep consolidates spaced memories through hippocampal replay. Dopamine tags hard-won information as valuable. Reconsolidation opens a window for each retrieved memory to be rebuilt stronger than before.

And yet the biggest obstacle remains the same one Bjork identified thirty years ago. The fluency illusion. The deep, persistent human tendency to mistake ease for effectiveness. To choose rereading over testing. To choose blocking over interleaving. To choose cramming over spacing. To choose comfort over growth.

The science is clear. The question is whether anyone will listen to it.