Our method
Re-reading and highlighting produce fluency, the comfortable sense that you know something. Fluency is not retention. Every mechanic in LearnForge is built around the small number of techniques that reliably survive controlled study, even though all of them feel worse in the moment.
Not as marketing language. As the actual control flow of the product. Each one maps to a specific behaviour you can observe while using it.
You are asked before you are told. Producing an answer, even a wrong one, creates a memory trace that reading the same sentence twice never will. Every Learn session opens with a question, not a summary.
Testing effect · Roediger & Karpicke
Sessions start just past the edge of what you've demonstrated, not at the beginning of the chapter. Too easy and nothing encodes; too hard and you disengage. The adaptive step targets roughly 80% success.
Zone of proximal development · Bjork
When you miss, the tutor names the specific misconception and asks a narrower question rather than reciting the answer. The explanation only arrives once you've committed to a position.
Guided discovery · Chi & VanLehn
Each card carries a stability and a difficulty. Reviews are scheduled for the moment your recall probability drops to 90%, late enough to be effortful, early enough to succeed.
Forgetting curve · Ebbinghaus, FSRS-style scheduling
Review queues mix concepts instead of blocking one chapter at a time. Discriminating between similar ideas is most of what an exam actually measures.
Interleaved practice · Rohrer & Taylor
Every session closes with a free-text prompt asking you to connect the pieces in your own words. Generating the explanation is the step that converts recognition into understanding.
Self-explanation effect · Chi et al.
Six stages, typically under two seconds for a chapter. The retrieval layer is the part most tools get lazy about, and it is the single biggest determinant of whether answers are trustworthy.
Your file is read in the browser. Layout-aware extraction rebuilds reading order from glyph coordinates, so two-column PDFs and lecture slides don't come out scrambled.
Text is split on semantic boundaries, headings, paragraphs, sentence groups, with overlap, so a definition is never cut in half. Each chunk keeps its page and heading.
A BM25 keyword index and a vector index are built side by side. Keyword search nails exact terminology; embeddings catch the paraphrase you actually typed.
Queries run against both, scores are fused, and MMR re-ranking removes near-duplicates so the model sees five distinct passages rather than the same paragraph five times.
The model is handed only those passages and instructed to cite them. If the answer isn't in your material it says so instead of inventing one.
What you got wrong becomes a card. The scheduler decides when you'll see it again. And the analytics page shows you the forecast.
The scheduler
Stability is how long the memory survives, the half-life of your recall. It grows each time you successfully retrieve the card, and it grows more when the retrieval was difficult.
Difficulty is how resistant this particular item is to becoming stable. It nudges up when you press Again and drifts down when you press Easy, so intrinsically hard material keeps getting the extra exposure it needs.
The next interval is the point at which predicted recall falls to 90%. You will feel slightly unsure at every review. That feeling is the mechanism working, not a bug in the schedule.
A card's life, in intervals
Five reviews spread over six weeks beat twenty-five reviews crammed into one night. And take a fifth of the time.