Bruner · Science
The science of making it stick.
Bruner is an AI tutor built on four findings from cognitive science — each one decades old, replicated, and mostly ignored by the tools people actually use to learn. Here they are, with the receipts.
Retrieval practice
You remember what you retrieve, not what you reread.
In a classic experiment (Roediger & Karpicke, Psychological Science 2006), students either restudied a passage repeatedly or studied it once and then recalled it from memory. Five minutes later, the rereaders looked better. A week later, the results had flipped: the students who practiced retrieval retained substantially more — and the rereaders’ advantage had evaporated.
Rereading and lecture-watching produce fluency: the material feels familiar, so you judge it learned. Retrieval is the thing that actually rewires memory.
This is why Bruner asks before it tells. Lessons run as dialogue — you answer, explain, and apply before the tutor fills the gaps — so every concept gets retrieved at the moment you’re building it, not just recognized on a slide.
Recall after a delay · restudying vs retrieval
Data: Roediger & Karpicke (2006), Expt. 2 — free recall of prose passages after repeated studying (83% → 40%) vs repeated retrieval practice (71% → 61%).
The science of remembering
And then it makes sure you don't forget.
Most apps make you cram. Bruner schedules each review at the exact moment you're most likely to forget — built from your lessons, calibrated to what you specifically got stuck on.
No cards to write, no schedule to maintain.
So what you understood in week 2 is still there in week 12 — and a year later when you actually need it on the job.
Retention over 30 days · With vs without spaced repetition
Baseline: Ebbinghaus's forgetting curve on novel material.
Desirable difficulty
Struggle is the feature. Frictionless help is debt.
The Bjork lab at UCLA spent decades showing that conditions which make practice feel harder — spacing, interleaving, testing — produce more durable learning than conditions that feel smooth (“desirable difficulties”).
Modern AI made the smooth path infinitely smoother — and measurably worse. In a 2025 Wharton/Penn study (Bastani et al., PNAS), students practicing with unrestricted ChatGPT looked 48% better during practice, then scored 17% worse on the unassisted exam. The MIT Media Lab (Kosmyna et al., 2025 preprint) called it “cognitive debt” — 83% of ChatGPT-assisted writers couldn’t quote from the essay they’d just written.
Bruner’s contract is the opposite of frictionless: it asks, hints, and reviews — and by design won’t do the work for you. The struggle isn’t a UX flaw to optimize away. It’s the mechanism.
Wharton / Penn 2025 · unrestricted-ChatGPT arm
Same students, two phases
Performance vs the no-AI control arm. Center line = control baseline. The lift was borrowed, not earned.
Backward design
Start from what you'll create. Teach backward from there.
Most courses are built forward from a syllabus: here’s the material, hope you find a use for it. Bruner builds backward from a capstone — the app, the analysis, the defended essay you actually want to exist. Every concept earns its place by being needed for the thing you’re building, which is also what makes retrieval practice meaningful instead of trivia.
And because an LLM that freestyles is a liability in education, Bruner grounds every lesson in research packs — current docs, papers, canonical books, and anything you upload — with citations inline. The model teaches from sources, not from vibes.
Built backward
The capstone
“A working AI agent, deployed, in my GitHub.”
Syllabus
Only the concepts the capstone actually needs.
Lessons
Taught as dialogue — you retrieve before the tutor tells.
Reviews
FSRS-scheduled, right before you'd forget.
Ship
The artifact is the exam. Defended in writing.
Is this actually research-based, or research-flavored marketing?
The four load-bearing claims — retrieval practice beats rereading, spacing beats cramming, productive struggle beats frictionless answers, and unstructured AI help can hurt exam performance — each link to the primary source right where the claim is made. We don't cite anything we haven't read, and where the evidence is thinner (transfer to every subject, for example) we say so below.
Why not just watch lectures? It feels like learning.
That feeling is the problem. Passive review produces fluency — the material feels familiar, so you judge it learned. Retrieval practice feels harder and slower, which is why almost nobody chooses it voluntarily. The research is blunt: the harder path is the one that survives a delay. Bruner's job is to make the harder path the default.
Does spaced repetition work for conceptual subjects, or just vocabulary?
The strongest effects were first shown on simple materials, but testing and spacing effects replicate on prose passages, science concepts, and problem-solving procedures. Bruner also doesn't rely on flashcards alone: exercise cards regenerate as fresh problems on the same concept, so a review is retrieval of the idea, not recognition of the card.
Where does the research not back Bruner's approach?
Honest limits: most studies measure weeks-to-months, not years. Hands-on motor skills and taste-based crafts benefit less from retrieval scheduling. And no study has tested Bruner itself yet — we're applying the best-supported general principles, not claiming a clinical trial of our product. When we run our own retention evaluations, we'll publish the method.
Learn it once. Keep it.
Fall 2026 private beta. Founder pricing locked for the life of your subscription.
23% off during beta · No credit card required