Knowly's Theory of Learning
Fifty years of cognitive science.
This is the research that shaped how we think about learning — and the foundation every decision in Knowly was built on.
3 Pillars
Knowly's Theory of Learning
The Learning Edge
Real growth only happens at the precise edge of what a child can almost — but not yet — do.
The Evidence
Vygotsky called it the zone of proximal development (1934): the narrow band where material is just hard enough to require effort but achievable enough to avoid shutdown. Below it, the child confirms what they already know. Above it, they disengage. In 2019, Wilson et al. published a mathematical proof in Nature Communications — the 85% Rule — demonstrating that learning rate peaks when a student succeeds on roughly 85% of first attempts. Independently, Rosenshine found that the most effective teachers maintained an 80–82% first-attempt success rate in guided practice; less effective teachers sat at 73%. This is not intuition. It is arithmetic. A teacher with twenty-two students cannot find and hold twenty-two moving edges simultaneously. The math does not allow it.
How Knowly Implements It
A diagnostic on day one places every child at their actual current level in every subject — not their grade label. From there, every response is analyzed for exactly where the reasoning broke down. Not "weak in algebra" — sets up the equation correctly, makes consistent sign errors when moving terms across the equals sign. That named gap becomes the child's current ceiling. The adaptive engine continuously targets a 70–85% first-attempt success rate. Too high — difficulty increases. Too low — the platform steps back and reteaches. The ceiling moves only when the work proves it has moved.
The Climb
Every rep has to be done by the person who needs to get stronger.
The Evidence
Learning is a physical event — synaptic connections literally strengthen through effortful thinking. This only happens when the child's own brain does the cognitive work. Slamecka and Graf (1978) demonstrated in five controlled experiments that self-generated content is encoded significantly more deeply than externally-provided content — the generation effect. Stockard et al. (2018) meta-analyzed 328 studies across fifty years and found that explicit instruction followed by student practice produced an effect size of d = 0.60, while unguided discovery produced d = −0.38. Guided practice is not optional scaffolding — it is the mechanism. Van der Weel and van der Meer (2024) used high-density EEG to show that handwriting recruits dramatically broader brain connectivity than typing — precisely the circuitry associated with memory formation. When a chatbot writes the essay or a calculator solves the equation, the child's brain undergoes none of the process that would have built real understanding. The shortcut economy is not a convenience. It is cognitive theft.
How Knowly Implements It
Knowly provides the instruction, the examples, and the scaffolding — then requires the child to generate the answer themselves, every time. Handwritten work is required across subjects: the child writes on paper, photographs it, and the platform analyzes the reasoning behind the answer, not just the final result. Voice recording captures reading fluency and oral work. Keystroke biometrics, time-on-task monitoring, and response pattern analysis ensure the work is genuinely the child's. The platform will not advance until understanding is demonstrated in multiple formats. There is no version of Knowly that does not require doing the actual work.
The Muscle
What is learned must be brought back — at the right moment, in a different form — or it disappears.
The Evidence
Ebbinghaus documented the forgetting curve in 1885: without reinforcement, roughly 67% of newly learned material is gone within a day. Murre and Dros replicated his findings precisely in 2015. Re-reading and re-studying feel like learning and produce almost none of it. The cure is spaced retrieval — the same concept returned at expanding intervals, just as it is about to slip below the retention threshold. Cepeda et al. (2006) meta-analyzed 184 studies and confirmed spaced practice consistently outperforms massed practice. Hattie and Donoghue (2021) analyzed 242 studies covering 169,179 participants and concluded that distributed practice and practice testing are the two most effective learning techniques known — above every other strategy studied. Rohrer et al. (2014) tested interleaved practice in 7th-grade math and found a 72% correct rate versus 38% for blocked practice: an effect size of d = 1.05 in a real classroom.
How Knowly Implements It
Every concept a child masters enters a per-child spaced retrieval schedule. It returns in days, then weeks, then months — calibrated to each child's modeled retention for that specific concept. When it returns, the platform does not show the answer and ask for re-reading. It presents a new problem and requires the child to generate the answer from memory. Returning concepts are interleaved with related concepts — the child must identify which strategy applies, not simply pattern-match to the current unit. The curriculum advances. The foundation never erodes.
2 Sigma
Gain with 1:1 mastery instruction
Bloom, 1984
85%
Optimal first-attempt success rate
Wilson et al., 2019
d = 1.05
Interleaving effect size in real classrooms
Rohrer et al., 2014
+150%
Retention from active recall vs. re-study
Karpicke & Roediger, 2008
Study in Focus
The mathematics of the learning edge.
In 2019, Robert Wilson and colleagues at Arizona, Brown, UCLA, and Princeton published a paper in Nature Communications that did something unusual in education research: it proved an optimal learning condition mathematically.
Using machine-learning algorithms working through classification tasks, they demonstrated that learning rate is maximized when the model succeeds on approximately 85% of attempts — an error rate of exactly 15%. Too easy (success rate above 85%) and the model stops improving. Too hard (success rate below 70%) and it fails to converge. The productive zone is narrow.
Wilson explicitly connected this finding to the human educational concept of the zone of proximal development. The authors noted that the 85% rule "provides quantitative support for the concept of the 'zone of proximal development.'"
Separately, Barak Rosenshine analyzed classroom observation data across decades and found that the most effective teachers maintained student success rates of 80–82% during guided practice. Less effective teachers ran at 73%. Neither knew the other's work. Both converged on the same zone. Knowly's adaptive engine targets 70–85% first-attempt success rate in real time — not as a rule of thumb, but as a mathematically grounded optimization.
The optimal learning zone
Above 85% success rate
Material is too easy. The child confirms existing knowledge but builds nothing new. Boredom and rushing follow.
70–85% success rate
The productive zone. Hard enough to require genuine effort. Achievable enough to maintain engagement. This is where real learning happens.
Below 70% success rate
Material is too hard. The child shuts down, disengages, or guesses randomly. No learning occurs and confidence erodes.
Wilson et al. (2019), Nature Communications. Independently corroborated by Rosenshine classroom observation data (2012).
A teacher with twenty-two students cannot track twenty-two moving edges in real time.
Knowly's adaptive engine recalibrates difficulty after every graded response — for every child, in every subject, every session.
The generation effect — what the research shows
Child reads the answer
Weak encoding. Feels like learning. Evaporates within days.
Child copies the answer
No encoding. Recognition without understanding.
Chatbot generates the answer
Zero encoding. The child's brain was not involved.
Child generates the answer
Deep encoding. Synaptic pathways strengthen. This is what learning is.
Slamecka & Graf (1978). Bjork (1994, 2011). Van der Weel & van der Meer (2024).
Study in Focus
Why the shortcut economy is a cognitive catastrophe.
In 1978, Norman Slamecka and Peter Graf at the University of Toronto published what became known as the generation effect. Across five experiments, subjects who generated target words from cues remembered them significantly better than subjects who simply read the same words — even with equal time on task. The act of generation, not the exposure, was what produced memory.
UCLA psychologist Robert Bjork extended this in 1994 with the concept of "desirable difficulties" — conditions that make learning feel harder in the moment but produce more durable knowledge. Spacing, interleaving, retrieval, reducing feedback: all of them feel counterproductive while they are working. The effort is not a side effect. It is the mechanism.
When a calculator solves the equation, when a chatbot writes the essay, when a parent types the answer — the child is not spared the difficulty. They are robbed of the only moment that would have built something. The brain changes its structure through the act of doing hard things. There is no substitute for that process and no shortcut through it.
This is why Knowly requires handwritten work, voice responses, and keystroke biometrics. Not as integrity theater — as a direct implementation of what the science says learning actually requires. We believe in physical writing because the research demands it.
Full Citations
Every study we build on — with the finding and the source.
2 Sigma
Average gain with 1:1 mastery instruction vs. conventional classroom
Students receiving one-to-one mastery-based instruction performed two standard deviations above students in conventional classrooms — placing the average tutored student at the 98th percentile of classroom-taught peers. Bloom called finding a scalable version of this "the 2 Sigma Problem." For forty years, no answer existed. Adaptive AI instruction is the first serious candidate.
Bloom, B.S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6), 4–16.
85%
Optimal first-attempt success rate for maximum learning speed
Using machine-learning models, Wilson and colleagues proved mathematically that learning rate is maximized when training accuracy sits at approximately 85% — an error rate of 15%. The principle generalizes to human learning and maps directly to Vygotsky's zone of proximal development. Too easy and the learner stagnates. Too hard and they disengage. The edge is narrow, measurable, and different for every child in every subject.
Wilson, R.C., Shenhav, A., Straccia, M., & Cohen, J.D. (2019). The Eighty Five Percent Rule for Optimal Learning. Nature Communications, 10, 4646.
d = 0.60
Effect size of direct instruction across 328 studies, 50 years
A landmark meta-analysis of 328 studies spanning five decades found an overall effect size of d = 0.60 for direct instruction — explicit teaching followed by guided student practice. By contrast, unguided discovery learning produced an effect of d = −0.38, meaning it actively underperforms doing nothing. The child must be taught, and then must do the work themselves. Both halves are required.
Stockard, J., Wood, T.W., Coughlin, C., & Khoury, C.R. (2018). The Effectiveness of Direct Instruction Curricula: A Meta-Analysis of a Half Century of Research. Review of Educational Research, 88(4), 479–507.
+150%
Better long-term retention from active recall vs. passive re-study
Students who practiced retrieval — generating answers from memory — retained 150% more material after one week than students who restudied the same content for the same amount of time. The act of generating the answer is what builds the memory trace. Re-reading feels productive and produces almost nothing. This finding is one of the most replicated in cognitive psychology.
Karpicke, J.D., & Roediger, H.L. (2008). The Critical Importance of Retrieval for Learning. Science, 319(5865), 966–968.
d = 1.05
Effect size for interleaved vs. blocked practice in 7th-grade math
Students who received interleaved practice — problems from multiple concept types mixed together — scored 72% correct on a delayed test versus 38% for students who practiced concepts in blocks. Effect size: d = 1.05. An extraordinarily large effect in real classroom research. Interleaving forces the child to identify which strategy applies, not just execute the one they practiced five minutes ago. This is what makes knowledge transferable.
Rohrer, D., Dedrick, R.F., & Burgess, K. (2014). The benefit of interleaved mathematics practice is not limited to superficially similar kinds of problems. Psychonomic Bulletin & Review, 21(5), 1323–1330.
0.57 SD
Average effect size of mastery learning across 108 controlled studies
A meta-analysis of 108 controlled evaluations found a mean effect size of 0.57 standard deviations for mastery learning programs — a substantial, consistent improvement across different ages, subjects, and teaching contexts. Mastery learning is not a philosophy. It is one of the most-replicated instructional findings in the literature.
Kulik, C.C., Kulik, J.A., & Bangert-Drowns, R.L. (1990). Effectiveness of mastery learning programs: A meta-analysis. Review of Educational Research, 60(2), 265–299.
67%
Of newly learned material forgotten within 24 hours without reinforcement
Ebbinghaus documented the forgetting curve in 1885. Murre and Dros replicated his findings faithfully in 2015, 130 years later: roughly 56% of newly learned material is lost within an hour, and 67% within a day, without deliberate reinforcement. This is not a learning failure — it is how the brain conserves resources. The only counter is spaced retrieval, timed to catch concepts just before they slip below the retention threshold.
Murre, J.M.J., & Dros, J. (2015). Replication and Analysis of Ebbinghaus' Forgetting Curve. PLOS ONE, 10(7), e0120644.
17
Evidence-based principles of effective instruction identified by Rosenshine
Rosenshine synthesized decades of classroom research into 17 principles of effective instruction. Every one points in the same direction: explicit teaching, frequent low-stakes checking for understanding, worked examples, and substantial guided practice before independent work. These principles are the structural backbone of how Knowly sequences every lesson — introduction, demonstration, scaffolded practice, independent generation.
Rosenshine, B. (2012). Principles of Instruction: Research-Based Strategies That All Teachers Should Know. American Educator, 36(1), 12–19.
An honest note about what we know and don't know.
The research above is strong, well-replicated, and genuinely the foundation of how Knowly is built — not window dressing. We cite it because every feature in the platform traces back to one of these findings.
What the research does not tell us: whether Knowly specifically will produce 2 Sigma outcomes for every child. Bloom's findings came from human tutors. The question of whether adaptive software can replicate those results is an empirical question we intend to study honestly — including when the results are not what we hoped.
We will publish outcome data from enrolled families starting in late 2026. We will not cherry-pick results. If Knowly is not producing the outcomes the research predicts, you will read about it here before you read it anywhere else.
See the research in practice.
Join the early access list for our August 2026 launch.