The Algorithm That Remembers: Why 140 Years of Memory Science Still Can't Make You Fluent

Here is a fact that should bother you: scientists have known since 1885 that spreading your study sessions over time dramatically outperforms cramming. Hermann Ebbinghaus demonstrated this with his pioneering memory experiments using nonsense syllables, measuring what he called "savings" — the time saved when relearning material after a delay (Ebbinghaus, H. (1885). Über das Gedächtnis…). One hundred and forty years later, a landmark meta-analysis by Cepeda and colleagues examined 839 assessments from 317 experiments and confirmed it across every retention interval tested, from less than one minute to more than 30 days (Cepeda, N. J., Vul, E., Rohrer, D., Wixted…). Spaced practice didn't just edge out massed practice. It was superior in 95.6% of the comparisons.

And yet.

Education apps — the very tools designed to deliver this technique to millions — have the lowest user retention rates of any mobile app category, at just 1.76% (Claude synthesis (2025). Comprehensive res…). Only 0.1% of Duolingo's half-billion users complete a course (Claude synthesis (2025). Comprehensive res…). The most rigorously proven learning technique in all of psychology has, by any practical measure, a catastrophic adoption problem.

In 1988, educational psychologist Frank Dempster published a paper with a title that doubles as an indictment: "The Spacing Effect: A Case Study in the Failure to Apply the Results of Psychological Research" (Dempster, F. N. (1988). The Spacing Effect…). He found that neither American classrooms nor textbooks systematically implemented spaced review — and that, remarkably, Soviet mathematics textbooks provided more distributed presentation of material than their American equivalents. Nearly four decades later, Lindsey and colleagues would argue that providing optimal spaced practice "is beyond what any teacher or student can reasonably arrange" without technological support (Dempster, F. N. (1988). The Spacing Effect…).

So technology stepped in. Spaced repetition software — Anki, SuperMemo, Duolingo, Memrise, and dozens of others — promised to solve the arrangement problem algorithmically. And in many ways, these tools are extraordinary. But the story of spaced repetition in the real world is not a story of triumph. It's a story of a profound mismatch between what science knows, what apps deliver, and what learners actually do. Understanding that mismatch is what this episode is about.

To understand why spacing works so reliably, you have to go small — all the way down to the proteins inside your neurons. The biological case rests on two molecular players that most learners have never heard of: CREB and MAPK.

CREB — cyclic AMP response element-binding protein — functions as a molecular switch that determines whether a learning experience produces long-term memory or merely a temporary impression (BrainFacts.org (2021). The Neuroscience Be…). Think of it as a gate that must be opened before lasting memories can be written. The critical insight comes from elegant experiments with fruit flies. When researchers gave Drosophila ten odor-shock pairings in rapid succession — the insect equivalent of cramming — the flies learned to avoid the odor for about three days (BrainFacts.org (2021). The Neuroscience Be…). But when those same ten pairings were spread out with 15-minute rest intervals between them, the flies avoided the odor for seven days or more. For an organism whose entire lifespan is roughly 50 days, that's the difference between a Post-it note and a tattoo.

The proof that CREB is truly the rate-limiting step came from genetic manipulation. When researchers engineered flies to overexpress CREB, suddenly massed training — cramming — produced long-term memory too (BrainFacts.org (2021). The Neuroscience Be…). The training hadn't changed. The molecular gate had simply been forced open. Under normal conditions, spacing is the only way to open it.

MAPK — mitogen-activated protein kinase — provides the timing mechanism that explains why those 15-minute gaps matter (Smolen, P., Zhang, Y., & Bhatt, D. K. (201…). In cellular studies, four spaced 3-minute depolarizations with 10-minute rest periods evoked persistent MAPK activation. But collapsing that same stimulation into a single 12-minute pulse failed to produce the same effect (Smolen, P., Zhang, Y., & Bhatt, D. K. (201…). MAPK creates a roughly 45-minute temporal window after an initial learning event during which a second exposure can generate long-term memory. Miss that window by cramming everything together, and the molecular machinery never engages.

At the systems level, neuroscience adds another layer. Your fast-learning hippocampus temporarily stores new memories and then gradually transfers them to the slow-learning neocortex over days to weeks, primarily during sleep (Wang, J. et al. (2025). Spaced learning in…). Sharp-wave ripples during sleep compress and replay information for cortical consolidation. This transfer cannot be rushed — it operates on its own biological timetable. Recent fMRI research has confirmed this process directly: spaced learning induced higher neural pattern similarity in default mode network subsystems during retrieval compared to massed learning, and critically, this neural integration in the dorsal-medial and medial-temporal subsystems predicted durable memory persisting to a one-month delay (Wang, J. et al. (2025). Spaced learning in…).

The molecular story resolves a question that learners intuitively ask: is spacing just a study tip, or is it something deeper? The answer is unambiguous. Spacing is a biological requirement. The proteins that write long-term memories operate on their own timelines, and no amount of willpower or concentration can override molecular kinetics.

If spacing is biologically non-negotiable, the natural next question is: how do you space optimally? This is where algorithms enter the picture — and where the story gets surprisingly anticlimactic.

The modern history of spaced repetition algorithms begins with Piotr Woźniak, a Polish researcher who in 1985 conducted a personal learning experiment that would eventually spawn SuperMemo, the first commercial spaced repetition software (Woźniak, P. (2025). The True History of Sp…). His initial algorithm, SM-0, used fixed intervals of 1, 2, 4, 8, 16, and 32 days — a simple doubling pattern derived from his own study data rather than any theoretical model (Woźniak, P. (2025). The True History of Sp…). By 1987, he had developed SM-2, which replaced fixed intervals with adaptive matrices adjusted by an "ease factor" that tracked individual item difficulty (SuperMemo (2025). SuperMemo Algorithm docu…). Correct answers lengthened intervals; incorrect answers shortened them.

SM-2 has proven, in the words of one analysis, "remarkably durable." Thirty-eight years after its creation, it remains the scheduling engine behind Anki and Mnemosyne — two of the most widely used flashcard applications in the world (Woźniak, P. (2025). The True History of Sp…). Woźniak continued developing increasingly sophisticated algorithms through SM-18, incorporating forgetting curves, stability matrices, and decades of user data. But independent validation of these later versions remains limited, with most evidence coming from SuperMemo's own internal benchmarks (SuperMemo (2025). SuperMemo Algorithm docu…).

The most significant algorithmic challenger to emerge in recent years is FSRS — the Free Spaced Repetition Scheduler — now integrated into Anki as of early 2025, with predictions it could become the default scheduler by late 2025 (Anki Forums / Reddit (March 2025). FSRS-5…). FSRS represents a genuine technical advance. It models memory through three components: retrievability (the probability you'll recall an item), stability (the interval at which retrievability drops to 90%), and difficulty. Its 21 trainable parameters are optimized via machine learning on individual user review histories (Ye, J. et al. (2025). FSRS algorithm speci…). A critical innovation: FSRS uses power-law forgetting curves rather than exponential ones, which provide a superior fit to observed data (Ye, J. et al. (2025). FSRS algorithm speci…).

In benchmarks across 727 million reviews from approximately 10,000 Anki users, FSRS achieved a log loss of 0.3460, compared to 0.4694 for Duolingo's Half-Life Regression algorithm — a substantial improvement in prediction accuracy (Ye, J. et al. (2025). FSRS algorithm speci…). Community sentiment has been largely positive, with users reporting reduced review burdens (Anki Forums / Reddit (March 2025). FSRS-5…).

But here's the twist that matters: prediction accuracy is not the same as learning outcomes. FSRS can tell you with greater precision when you're about to forget something. What it has not yet demonstrated is that this precision translates into meaningfully better retention over months or years compared to simpler algorithms (Claude synthesis (2025). Comprehensive res…). No rigorous head-to-head trials have compared long-term proficiency outcomes between FSRS and SM-2.

The meta-analytic evidence puts this in perspective. Expanding spacing schedules — the kind that sophisticated algorithms produce — outperform fixed spacing schedules by roughly 3% (Cepeda, N. J., Vul, E., Rohrer, D., Wixted…). Three percent. Meanwhile, any reasonable spaced algorithm outperforms massed practice by enormous margins. The implication is uncomfortable for algorithm enthusiasts: the vast majority of the benefit comes from spacing at all, not from spacing optimally. Cepeda and colleagues found that for one-week retention, optimal gaps fall between 20–40% of the retention interval; for one-year retention, 5–10% (Cepeda, N. J., Vul, E., Rohrer, D., Wixted…). Most commercial apps don't even ask what your retention goal is.

This doesn't mean algorithmic progress is meaningless. For medical students reviewing thousands of cards over years, a 20–30% reduction in unnecessary reviews — which FSRS may deliver — is genuinely valuable. But for most learners, the algorithm is not the bottleneck. The bottleneck is everything else.

There's a phenomenon that language-learning communities describe with a mixture of frustration and dark humor: the learner who has reviewed 20,000 Anki cards and cannot hold a basic conversation. It's not an edge case. It's the predictable outcome of a fundamental gap in how most spaced repetition systems work.

The Kim and Webb 2022 meta-analysis — 48 experiments, 3,411 participants — confirmed that spaced practice produces large effect sizes for vocabulary learning (Kim, S. K. & Webb, S. (2022). Meta-analysi…). But the authors included a crucial caveat: "the majority of studies focus on paired-associate learning" and measure outcomes "in formats similar to how material was learned" (Kim, S. K. & Webb, S. (2022). Meta-analysi…). In other words, the studies proved that flashcard users get better at flashcards.

The problem is that recognition and production appear to be fundamentally different cognitive processes. González-Fernández's 2025 study of 314 EFL learners found that recognition knowledge precedes recall knowledge across all vocabulary components in a predictable developmental sequence (González-Fernández, B. (2025). Recognition…). Stewart and colleagues went further in 2024, arguing that lexical recall and recognition may be "distinct psychometric constructs" — different enough to function as separate abilities rather than points on a single continuum (Stewart, J. et al. (2024). Lexical recall…).

The practical consequences are severe. Research has found that vocabulary knowledge explains 32–84% of speaking proficiency variance depending on conditions, but — and this is the critical finding — "learners with large vocabulary sizes did not necessarily produce lexically sophisticated L2 words during speech" (Claude synthesis (2025). Comprehensive res…). Recognition creates what researchers call an illusion of knowledge that production exposes as shallow.

Why does this happen? Several well-established theoretical frameworks converge on the same answer. DeKeyser's skill acquisition theory holds that the declarative knowledge SRS builds — knowing what a word means — must transform into proceduralized knowledge through production practice over many trials before it becomes available for spontaneous use (Claude synthesis (2025). Comprehensive res…). Flashcard review is controlled, deliberate processing; spontaneous speaking requires automatic processing. These are different neural pathways.

Then there's transfer-appropriate processing: memory works best when encoding conditions match retrieval conditions. Reading a Japanese character on a white Anki card in your bedroom engages fundamentally different neural processes than hearing that word in a noisy izakaya and needing to respond in 400 milliseconds (Claude synthesis (2025). Comprehensive res…). And context-dependent memory — demonstrated dramatically by Godden and Baddeley's classic study showing that words learned underwater were recalled better underwater (mean 24.9) than on land (mean 17) — suggests that the interface itself becomes part of the memory trace (Claude synthesis (2025). Comprehensive res…).

Finally, SRS provides no communicative pressure. Real conversation demands real-time lexical access under the stress of formulating a message while someone waits for your response. Flashcard review, by contrast, is self-paced, low-stakes, and binary. The gap between these two experiences is not a minor detail; it's the central reason why flashcard fluency doesn't transfer to conversational fluency.

None of this means SRS is useless for language learning. It means it's incomplete. And the difference between those two things matters enormously for how you spend your study time.

Let's talk about the elephant in the room: the companies building spaced repetition tools don't always have the same goals as the people using them.

Duolingo is the dominant player in language-learning technology, with over 500 million total users and 103.6 million monthly active users (Duolingo company metrics (2024–2025). 500M…). But only about 2% convert to paid subscribers, which means the company's revenue depends heavily on engagement metrics — daily active users, session length, streak maintenance — that keep eyeballs on ads and free users moving toward conversion (Duolingo company metrics (2024–2025). 500M…). Users who maintain a 7-day streak are 3.6 times more engaged than those who don't, which explains why streak mechanics dominate the user experience (Duolingo company metrics (2024–2025). 500M…).

The problem is that optimizing for engagement and optimizing for learning are not the same thing. A 2021 systematic review published in Taylor & Francis painted what the authors called "a mixed (and sometimes negatively skewed) picture" of Duolingo's effectiveness (Systematic review of Duolingo effectivenes…). The review concluded that the app's design decisions prioritize "competition over collaboration, repetition and translation over meaningful feedback and context, and passive receptive skills over active productive skills" (Systematic review of Duolingo effectivenes…). Once the novelty of gamification wore off, the authors argued, it could not compensate for these structural limitations.

The conflict is structural, not incidental. Engagement metrics — DAU, session frequency, time-on-app — are easy to measure and directly drive revenue. Learning outcomes — delayed recall, transfer to conversation, writing accuracy — are expensive to measure and may actually require shorter, less frequent sessions than engagement metrics reward (Claude synthesis (2025). Comprehensive res…). The heart system monetizes mistakes by requiring users to purchase hearts or watch ads to continue practicing. Push notifications are optimized by multi-armed bandit algorithms for maximum click-through rates, not for optimal learning timing (Duolingo company metrics (2024–2025). 500M…).

Eight years after research on Duolingo began in earnest, the systematic review noted that "we still have very little conclusive evidence about its effectiveness" (Systematic review of Duolingo effectivenes…). For a product used by over half a billion people, that's a striking gap.

Anki occupies the opposite end of the spectrum. It's open-source, user-owned, and treats itself as a toolkit rather than a curriculum (Anki Forums — Collection of Anki Resources…). The active add-on ecosystem — AnkiAIUtils, custom schedulers, elaborate template systems — reflects a design philosophy that prioritizes user control and scheduling transparency over guided simplicity (Anki Forums — Collection of Anki Resources…). FSRS-5 was adopted through community discussion and iterative testing, not a corporate product roadmap (Anki Forums / Reddit (March 2025). FSRS-5…). The trade-off is real: Anki's learning curve is steep, its interface is utilitarian, and it shifts the burden of card quality and study design entirely to the user.

Memrise has tried to split the difference, but its 2024–2025 pivot illustrates the tension. A "new experience" rollout in July 2025 emphasized immersive personalization, while community-created courses — the content that many users originally came for — were relocated to a separate site (Grok synthesis (2025). Real-time survey of…). Forum sentiment was mixed: relief that community content survived, frustration at the fragmentation. A Memrise-to-Anki migration thread on the Anki forums accumulated 102 replies and 8,594 views, signaling meaningful user demand for content portability when platforms change direction beneath them (Anki Forums — Collection of Anki Resources…).

The broader lesson is that SRS apps exist in a market where the incentives of the builder and the needs of the learner are imperfectly aligned. Engagement is measurable, monetizable, and optimizable. Learning is none of those things at scale. Users who understand this misalignment can navigate it; those who don't may mistake streak maintenance for actual progress.

Even if every app were perfectly designed and every algorithm flawlessly calibrated, spaced repetition would still face a fundamental obstacle: it feels terrible.

This isn't a minor UX complaint. It's a well-documented cognitive illusion. In one study, 83% of participants rated massed practice as equally or more effective than spaced practice — despite spaced practice producing objectively superior retention on delayed tests (Kornell, N. & Bjork, R. A. (2008). Learnin…). Learners consistently, reliably, and confidently prefer the method that works worse.

The mechanism is what psychologists call a fluency heuristic (Kornell, N. & Bjork, R. A. (2008). Learnin…). When you cram, material remains fresh in working memory. Retrieval feels smooth and effortless. Your brain interprets this fluency as evidence of strong learning. When you space your practice, you return to material after a delay. Retrieval is effortful, halting, uncertain. Your brain interprets this difficulty as evidence that the method isn't working (Hendrick, C. (2025). What Makes Spaced Pra…). The subjective experience is exactly backwards: the struggle that signals effective long-term encoding feels like failure.

This misalignment between feeling and reality creates what researchers call the judgments-of-learning paradox (Dempster, F. N. (1988). The Spacing Effect…). Students show a clear preference for massed repetition when judging learning effectiveness, even when objective tests prove spaced practice superior. Spaced items feel "more detached from short-term memory... less effective" (Dempster, F. N. (1988). The Spacing Effect…). The implication for SRS users is direct: the days when your review sessions feel hardest — when cards you thought you knew slip away and your accuracy drops — are likely the days when the most learning is occurring.

Recent research has added an important nuance to why this happens. Two experiments comparing massed and spaced calculus learning administered working memory tests after each condition and found that working memory was not significantly depleted in either condition (Hendrick, C. (2025). What Makes Spaced Pra…). The old "rest and recovery" theory — that spacing works because your brain needs a break — doesn't hold up. Instead, evidence points toward mental rehearsal: even when you're not consciously thinking about the material, your brain continues processing it during the gaps between study sessions (Hendrick, C. (2025). What Makes Spaced Pra…). But this unconscious processing depends on having enough foundational knowledge to rehearse meaningfully, which may explain why spacing benefits increase with expertise.

The metacognitive illusion also explains the review-burden dropout spiral. When learners skip a day of Anki, they return to a growing pile: Day 1 leaves approximately 50 overdue reviews, Day 2 grows to 120, Day 3 to 190, Day 4 to 280 (Claude synthesis (2025). Comprehensive res…). Facing that mountain, the retrieval experience feels overwhelmingly difficult. The brain's fluency heuristic screams that this isn't working. And so the learner quits — not because the system failed, but because it felt like it did.

The most common mistake new SRS users make is learning too many new cards per day, which drives the review pile into unsustainable territory within weeks (Claude synthesis (2025). Comprehensive res…). The recommended calibration — 10 to 20 new cards daily, completing all due reviews before adding new material, sessions capped at 15 to 30 minutes — sounds modest precisely because it is (Claude synthesis (2025). Comprehensive res…). Users who survive three months of consistent practice are four times more likely to achieve their language goals. But reaching that three-month threshold requires tolerating a daily experience that your own metacognition insists is ineffective.

If spaced repetition alone can't produce fluency, and apps may not be optimizing for your learning, what does an effective system actually look like? The best evidence we have comes from two sources: polyglot practitioners and a handful of well-designed studies. Neither is perfect, but together they converge on a surprisingly consistent picture.

Steve Kaufmann, founder of LingQ and speaker of 20+ languages, frames SRS as strictly secondary: "If you like doing flash cards, using spaced repetition systems, then it's worth doing. If not, this kind of learning activity won't help much" (Polyglot practitioner testimony — Steve Ka…). His emphasis falls on massive amounts of comprehensible input — listening and reading. Luca Lampariello, who has learned 20 languages, reports using SRS "only for a few specific needs" and prefers repeated exposure in context (Polyglot practitioner testimony — Steve Ka…). On the other end, Gabriel Wyner's Fluent Forever method positions SRS as central, but with important modifications: learn pronunciation first, avoid translations, and create cards that connect multiple information chunks — spelling, pronunciation, image, personal association, and grammatical gender (Polyglot practitioner testimony — Steve Ka…).

Despite their divergent prescriptions, these practitioners agree on core principles: SRS supplements but never replaces authentic interaction; personally created cards substantially outperform pre-made decks; daily consistency matters more than session length; and excessive SRS leads to burnout (Polyglot practitioner testimony — Steve Ka…).

The Refold methodology, which emerged from online language-learning communities, suggests beginners allocate 30–40% of study time to SRS, intermediates 20–30%, and advanced learners 10–15% or less (Polyglot practitioner testimony — Steve Ka…). These ratios are practitioner-derived heuristics, not the output of controlled trials — the research on optimal time allocation is, as one synthesis put it, "frustratingly sparse" (Claude synthesis (2025). Comprehensive res…). But they align with a theoretical model that resolves the apparent conflict between SRS advocates and immersion advocates: SRS builds the vocabulary floor needed to understand input, while comprehensible input provides the rich contextual exposure needed for acquisition (Claude synthesis (2025). Comprehensive res…).

A meta-analysis of 21 extensive reading studies (N=1,268) found effect sizes of d=1.32 for vocabulary gains from reading — comparable to SRS effect sizes (Claude synthesis (2025). Comprehensive res…). This suggests that for learners past the absolute beginner stage, extensive reading may be as powerful as flashcard review for vocabulary building, while simultaneously providing the context, grammar exposure, and processing practice that flashcards cannot.

For card design, the evidence points toward several evidence-backed strategies. Sentence cards teach vocabulary and grammar simultaneously, showing words in natural context (Claude synthesis (2025). Comprehensive res…). The "1T sentence" principle — only creating cards from sentences where you understand everything except one target element — ensures cards remain comprehensible and personally relevant (Claude synthesis (2025). Comprehensive res…). Dual-coding approaches, drawing on Paivio's finding that activating both verbal and visual processing facilitates retention, consistently outperform text-only cards, and self-generated mnemonics outperform provided ones (Claude synthesis (2025). Comprehensive res…). So-called "anime cards" — a target word highlighted within a sentence context, often with audio — can be reviewed 2–4 times faster than full sentence cards while preserving contextual benefits (Claude synthesis (2025). Comprehensive res…).

The metaphor that best captures the integrated approach: "When you make a flashcard out of something, it's like you get a cup. As you interact with your target language, you fill that cup with water" (Claude synthesis (2025). Comprehensive res…). SRS creates the containers. Everything else fills them.

The spaced repetition landscape is changing faster in 2024–2025 than at any point since Woźniak wrote SM-2 in 1987. Three developments deserve attention — and one persistent problem deserves honesty.

First, AI-assisted card generation is crossing the adoption threshold. Tools like AnkiAIUtils add AI-generated explanations, mnemonics, and images to existing cards. Template integrations with GPT allow users to generate contextually rich flashcards from PDFs, textbooks, and web content (Anki Forums — Collection of Anki Resources…). A survey cited on Anki forums found that 53% of medical students would use ChatGPT to generate Anki cards if tutorials existed — suggesting the barrier to adoption is knowledge distribution and workflow packaging, not AI capability (Anki Forums — Collection of Anki Resources…). Early comparisons show GPT-4 outperforming offline LLMs for card generation quality, though community caution about AI-generated cards introducing errors or "bad habits if unchecked" is well-placed (Grok synthesis (2025). Real-time survey of…).

Second, FSRS-5's integration into Anki represents the most significant scheduling upgrade the platform has seen in years. Community adoption has been largely positive, with users reporting improved efficiency and the algorithm predicted to become Anki's default by late 2025 (Anki Forums / Reddit (March 2025). FSRS-5…). The broader ecosystem is also maturing: tools like AnkiPandas allow programmatic analysis of collection data, enabling learners to audit their own forgetting patterns and adjust strategies accordingly (Anki Forums — Collection of Anki Resources…).

Third, guided platforms are investing heavily in features that may address some of the recognition–production gap. Duolingo's AI Video Calls and Adventures (September 2024) introduce interactive practice formats that go beyond flashcard-style recognition (Grok synthesis (2025). Real-time survey of…). Its September 2025 updates added PvP modes and LinkedIn integrations for professional application (Grok synthesis (2025). Real-time survey of…). Whether these features produce meaningful proficiency gains or primarily serve engagement metrics remains to be seen.

But the honest assessment is that none of these developments address the deepest problems the research identifies. The metacognitive illusion — that spacing feels worse than cramming — isn't solvable with better algorithms. The recognition–production gap isn't solvable with better flashcards. The 140-year adoption failure in formal education isn't solvable with better apps. And the structural conflict between engagement-driven business models and evidence-based learning design persists regardless of which AI model generates the cards.

The research reveals a technology that is simultaneously one of the most proven interventions in cognitive science and one of the most misunderstood by its users. Spaced repetition works. It works for reasons we can trace down to individual proteins. Modern algorithms have made it more efficient. But the gap between what the science offers and what learners achieve remains vast — not because the tools are broken, but because the tools were only ever meant to be one part of a larger system. The learners who succeed are the ones who build that system. And the ones who struggle are often the ones who mistake the tool for the whole.