Beating the Forgetting Curve: Spaced Repetition and Notes Studio

Every student knows the ritual. The exam is tomorrow, the notes are a week old and untouched, and the night dissolves into re-reading the same three pages until the words stop meaning anything. It feels like work. It even feels like progress. And then, two weeks later, almost none of it is left. Cramming is one of the most reliable ways to convert effort into the illusion of learning.

At pdpspectra we build learning and operations platforms for an international audience, and the memory-science layer inside PDP Shikshya, our EdTech product, exists precisely because cramming fails so predictably. This post is the engineering walkthrough of that layer: spaced-repetition flashcards driven by the SM-2 algorithm, and a Notes Studio that turns a student’s own notes into flashcards, a quiz, and a summary — removing the one piece of friction that stops most people from ever spacing their practice at all.

If you want the wider context first, start with the platform overview and the gamified learning post.

Why cramming loses to the forgetting curve#

More than a century ago Hermann Ebbinghaus measured how memories decay, and the shape he found — the forgetting curve — has held up ever since: newly learned material drops away fast at first, then more slowly, unless something intervenes. Cramming pours everything in at once, right before the peak of that curve, and then lets it all slide off together.

Two of the most robust findings in cognitive psychology point the other way:

• The spacing effect — the same total study time produces far more durable memory when it is distributed across days than when it is massed into one session. Spreading the work is not just gentler; it is measurably better retention for the same minutes spent.
• The testing effect — actively recalling an answer strengthens memory more than passively re-reading it. The struggle to retrieve is the learning event, which is why a flashcard beats a highlighter.

Put those together and the optimal strategy is almost the opposite of cramming: short bursts of active recall, spaced out over time, with the gaps getting longer as the memory gets stronger. The problem is that no student wants to hand-manage a review calendar for two hundred facts across five subjects. That scheduling is exactly the kind of tedious bookkeeping software should own.

SM-2: scheduling reviews at the point of forgetting#

The flashcard engine at /app/review runs the SM-2 algorithm — the classic SuperMemo-2 scheduling logic that has quietly powered serious spaced-repetition tools for decades. The idea is elegant and worth understanding, because it is the whole reason the system works.

Every card carries two numbers: an ease factor and an interval. When a card comes up for review, the student grades how well they recalled it. The algorithm then updates both numbers:

• A confident, correct recall stretches the interval. A card you have nailed three times might not reappear for weeks, then months — the system stops wasting your time on things you already know.
• A hesitant recall grows the interval more cautiously and nudges the ease factor down, so a shaky card comes back sooner.
• A lapse — a card you got wrong — resets the interval to the start. The card re-enters short-term rotation and has to earn its way back out.

The effect is that each card is scheduled to resurface roughly at the moment you are about to forget it. Review too early and you waste effort on a memory that is still fresh; review too late and you have already lost it and have to relearn from scratch. SM-2 targets the sweet spot in between, and it does so per card, so easy facts fade into the background while stubborn ones get the attention they need. The scheduling lives in srs.py, the card model and deck logic in flashcards.py.

Why we chose the boring, proven algorithm#

There is a whole zoo of newer scheduling algorithms, and it would have been easy to reach for the most fashionable one. We deliberately did not. SM-2 is decades old, thoroughly understood, and its behaviour is predictable enough that a teacher or a student can build an accurate mental model of it: get it right, see it less; get it wrong, see it soon. That legibility matters more in a classroom tool than a marginal efficiency gain from a black-box scheduler. When a student asks “why is this card back already?”, the honest answer — “because you missed it last time” — is one they can act on. Proven and explainable beat novel and opaque when children’s study time is the thing being optimised.

Notes Studio: closing the friction gap#

Spaced repetition has one notorious weakness, and it is not the algorithm — it is card creation. Making good flashcards is slow, finicky work, and it is the step where most students quit before they start. A review system that assumes a full, well-authored deck is a system most people never actually use.

Notes Studio at /app/notes is our answer to that. A student pastes or writes their own notes — a page copied from their exercise book, a messy paragraph from class, a chapter summary in their own words — and the AI generates three things from that exact material:

• Flashcards — question-and-answer pairs pulled straight from the student’s notes, ready to drop into the SM-2 review loop.
• A quiz — practice questions that turn passive notes into active retrieval, invoking the testing effect on the student’s own content.
• A summary — a condensed version that surfaces the core ideas and helps a student check whether their notes actually captured the point.

The critical design choice is that everything is generated from the student’s own notes, not from a generic textbook. The cards speak in the same terms the student wrote, cover the exact scope their teacher taught, and match the way the material was framed in their classroom. That grounding is what makes the output feel like their study material rather than a stranger’s. And because the flashcards flow directly into the SM-2 scheduler, the tedious step — building the deck — collapses from an evening of transcription into a few seconds. Note-taking and spaced practice stop being two separate chores and become one continuous loop. The generation logic lives in notes.py.

Privacy is not an afterthought here#

Because Notes Studio ingests whatever a student writes, and student notes can contain names, roll numbers, and other personal details, the privacy handling has to be strict. All AI in PDP Shikshya runs behind a swappable LLM client, and before any text reaches the model, student PII is stripped on-device by a local anonymizer — a combination of regex patterns and the class roster — so identifying details never leave the device in the first place.

Two things follow from that architecture:

• The model never sees who wrote the notes. Anonymization happens locally, ahead of the network call, so the text that travels is scrubbed of the details that could identify a child.
• The model is a replaceable component. Because everything sits behind one client interface, the underlying model can be swapped without touching the flashcard, quiz, or scheduling logic — a decision that keeps the platform durable as the AI landscape shifts, and keeps a school from being locked to a single vendor.

That is the same discipline we bring to every AI implementation at pdpspectra: treat the model as an interchangeable part, and put the guardrails in the code around it rather than trusting the model to behave.

Built for the curriculum it serves#

None of this is generic study software wearing a Nepal sticker. The whole platform is built for the local curriculum — CDC standards for grades 9 and 10, NEB for grades 11 and 12, and a broader grade 8 to 12 span in Learn mode — and it is fully bilingual across English and नेपाली. A student can take notes in Nepali and get Nepali flashcards back; the memory layer does not privilege one language over the other.

That curriculum-awareness compounds with spaced repetition in a way generic tools cannot match. Because the flashcards, quizzes, and summaries are scoped to the syllabus a student is actually sitting for, the reviews SM-2 schedules are reviews of exam-relevant material — not trivia, not off-syllabus facts, but the specific concepts the NEB or CDC expects them to hold. The memory science and the curriculum grounding reinforce each other: the right facts, surfaced at the right time, in the right language.

Why we built it this way#

Cramming is not a character flaw; it is a rational response to a system that never taught anyone how to schedule their own memory. The fix is not to nag students to “study earlier” — it is to make the correct strategy the path of least resistance. SM-2 removes the scheduling burden by deciding what to review and when. Notes Studio removes the authoring burden by building the deck straight from a student’s own notes. Together they turn the two hardest parts of spaced repetition — planning and preparation — into a background process.

That is the philosophy behind the whole build at pdpspectra: AI implementation and Data Platforms designed around what the operation actually needs, not around what demos well. The memory layer sits inside the same multi-tenant School ERP that runs attendance, homework, routines, quizzes, and the Socratic tutor — one platform where studying well is the default, not a discipline a student has to summon alone at midnight.

---

Want a learning platform where remembering is engineered, not left to willpower? See it at pdpshikshya.com, or talk to us about your build at pdpspectra.com. AI-powered. Data-driven. Built to ship.