The claims on the site are cited inline. This page holds the full references: what each one found, and which part of Ember stands on it.
Henry L. Roediger III & Jeffrey D. Karpicke (2006) · Psychological Science
Being tested on material produces more durable learning than studying it again, even when the restudy time is equal. The field calls this retrieval practice, and it is one of the most replicated effects in cognitive psychology.
On the site: practice exams that show their workJohn Dunlosky, Katherine Rawson, Elizabeth Marsh, Mitchell Nathan & Daniel Willingham (2013) · Psychological Science in the Public Interest
A review of ten study techniques. Practice testing and distributed practice earned the highest utility ratings; rereading and highlighting, the techniques students default to, earned the lowest.
On the site: flashcards that come back until they stickShen & Tamkin (2026) · Anthropic · arXiv:2601.20245
A randomized trial of engineers learning an unfamiliar library. AI users scored 17% lower on comprehension and debugging, and the gap was driven by interaction pattern: wholesale delegation produced the worst learning, while asking follow-up questions and requesting explanations preserved it. Ember's grounded chat is built for the second pattern.
On the site: ask questions inside the readingKosmyna et al. (2025) · MIT Media Lab · arXiv:2506.08872
An EEG study of essay writing with an LLM, a search engine, or neither. The LLM group showed the weakest neural engagement and largely could not quote from essays they had just submitted. The work the tutor leaves with the student is the work that builds anything.
Background: why the tutor asks instead of completingRAND Corporation · RRA134 series
The average teacher runs two core curricula plus five supplemental resources; 51% of high-school teachers teach primarily from self-created materials, and only 7% follow a single curriculum by the book. Teachers assemble; a lesson tool competes with their drive folder, not their publisher.
On the site: drop in your files, get a lesson backEmber (July 2026) · interviews and primary-source synthesis
Our own research on how teachers plan, deliver, and grade, verified against primary sources including Applebee & Langer (2011) and a 2025 CEP survey (n=937). The finding the site cites: written feedback returned after the fact mostly goes unread, with 81% of students spending fifteen minutes or less with it, and a visible grade suppressing attention to comments. Insight has to arrive while the next lesson can still change.
On the site: who needs help, and with exactly whatOpen-source skills and evaluation rubrics for K-12 teachers, co-developed with Learning Commons
The teaching skills behind Claude for Teachers: lesson planning and lesson differentiation as evaluated, versioned artifacts. A reference point for what teacher-grade AI tooling looks like when it is built with educators and published with its evals.
Reference: lesson building held to a published standard