Every brain
forgets differently.
Every classroom
teaches the same way.

Hermann Ebbinghaus  ·  1885
Forgetting Curve, With & Without Spaced Repetition
100% 75% 50% 25% 0% Retention Day 1 Day 3 Day 7 Day 14 Time No review Spaced retrieval

Memory decays. Not randomly, along a precise exponential curve understood for over a century. Almost no classroom has changed in response. Cognivia treats learning as something you can measure: an instrument that reads how each mind forms, holds, and loses knowledge, then adapts to what it finds. Proven first in India's hardest classrooms, and built to generalise.

What this work is for

The world already knows marks are broken.
The question is what replaces them.

Percentages tell you what was right on a given day. Percentiles tell you where you ranked against others on that same day. Neither tells you anything about how memory forms, how knowledge consolidates, or how a student thinks when no one is watching. The Cognitive Passport is a first attempt at answering those questions with evidence gathered continuously as a student works, not a single impression at term's end. It won't replace every form of assessment, but it may be a start at treating learning as the individual, biological process it is.

The Old Document

"What did you score?"

A number. A rank. An average of averages, designed to sort, not to understand. It compresses everything you are as a learner into a signal optimised for comparison, not growth.

The Cognitive Passport

"How does your mind actually work?"

A profile built from your own session data, that evolves with you and could one day tell a university, an employer, or a scholarship committee something no transcript ever could: the shape of your cognition.

Cognivia · Live session demo
Free to try, no sign-in needed
A full session, start to finish

Watch it measure how your memory works, in one session.

Five real questions sampled from Cognivia's science question bank. The engine classifies each wrong response by error type, updates your Learning Genome (λ RS, RLI, accuracy), and schedules every concept for review on a different day depending on how you answered.

At the end you can download a printable session report, same format the live trial generates.

≈ 3 minutes · 5 questions
Beyond Spaced Repetition

Most tools decide when to review.
Cognivia understands why you forget.

Getting an answer right or wrong is the shallowest thing about learning. Every session, Cognivia reads how your specific mind is working, and adapts to it.

01

Your mistakes, not just your score

A wrong answer is never simply wrong. Cognivia tells apart forgetting, a half-formed grasp, and a confident misconception, and responds to each one differently.

02

A forgetting curve for every concept

Each idea gets its own curve. The system learns how fast you personally lose it, then reschedules that idea on its own timeline, instead of one fixed interval applied to everyone.

03

The confidence that fails you on exam day

Answer with conviction and still get it wrong, and Cognivia flags it, surfacing the illusions of knowing that quietly fall apart under pressure.

04

Fluent mastery, or a lucky guess?

By watching how quickly you recall and how stable that memory becomes, it separates knowledge that will last from an answer you happened to get right.

05

Your Cognitive Passport

Everything above becomes a portable profile of how your mind learns, one that travels with the student, where a single score never could.

06

Runs anywhere, needs nothing

In any browser, on a shared classroom computer or a basic phone, with no install and little bandwidth. Adaptive learning without the infrastructure it usually demands.

The Core Problem

Students don't forget
because they're
not trying.

Over a century ago, Hermann Ebbinghaus demonstrated what is now called the Forgetting Curve: without deliberate reinforcement, most newly learned information disappears within days. This is not new knowledge. What is surprising is how little it has changed classrooms.

In most learning environments, especially those with limited resources, the structure remains unchanged: information is delivered, briefly practiced, tested, and then forgotten. The system assumes learning has happened, but memory rarely agrees.

Cognivia began with a single question: what if we designed learning around how memory actually consolidates, not around how teaching is conveniently organised?

EBBINGHAUS 1885 · MURRE & DROS 2015 · ROEDIGER & KARPICKE 2006
Forgetting without review
70% gone within 24 hours
Of newly learned material is forgotten within one day without structured reinforcement. Within a week the figure exceeds 90%.
Retrieval vs re-reading
+50% more retained at one week
Active retrieval practice outperforms passive re-study by this margin at delayed tests, even when total study time is held constant.
Years of research, classrooms unchanged
140 years since Ebbinghaus, no schedule change
The cognitive science behind spaced retrieval has existed since 1885. Standard classroom revision in most countries still does not use it.
What this work is for

The world teaches roughly 1.5 billion students every year. A tiny fraction reach anything resembling adaptive instruction. The classroom is not going to be rebuilt around each individual learner, not in any country, not in any decade soon. What can be rebuilt is the small block of revision each learner does on their own. Cognivia takes one narrow, falsifiable question: does adapting the schedule to the type of error a learner makes improve retention? Because if the answer is yes, the answer is portable: it scales to a shared device, it does not need a school to approve it.

If the trial comes back null, the repository and the data stay public, so the next person with a better hypothesis can start from there.

The Research Foundation

Two things science has known for decades.
Almost no classroom uses either.

Memory is not permanent by default. Newly learned information begins fading within hours unless it is revisited at precisely the right moments. A century of cognitive science has documented exactly how and when this happens. Almost none of it has reached the classrooms where it could make the most difference. That gap is what this work is about.

Mechanism 1

Spaced Practice

Distributing practice over increasing intervals exploits the spacing effect. Each successful retrieval strengthens the memory trace and extends the optimal review window. Cognivia implements the full adaptive review algorithm: a personalised ease factor and interval per card per student, updated after every response using a 1–5 quality rating. Struggling items return the next day; mastered items in weeks.

The review algorithm ease factor adjusts continuously per student per card, creating a schedule that is individualized, not uniform.

Effect size: high · d ≈ 0.6–0.8 across populations Lab → field test: rural classrooms, shared devices, no prior exposure
Ebbinghaus (1885) · Wozniak (1990) · Cepeda et al. (2006)
Mechanism 2

Active Recall

Forcing reconstruction of an answer from memory, rather than passive re-exposure, produces substantially deeper encoding. Roediger and Karpicke (2006) found retrieval practice outperformed re-study by ~50% on one-week delayed tests, even when re-study groups spent more total time. The generation effort is the encoding event.

Every Cognivia session is structured retrieval. The correct answer is revealed only after the student attempts a response. No passive exposure mode exists.

Effect size: high · 50% delayed recall advantage over re-study Lab → field test: cued retrieval practice in resource-constrained settings
Roediger & Karpicke (2006) · Bjork (1994) · Dunlosky et al. (2013)
Research Framing

Desirable Difficulty

There is a well-documented gap between what feels like effective studying and what actually produces lasting memory. Re-reading notes feels productive because the material looks familiar. But familiarity is not the same as genuine recall. The methods that feel harder in the moment, trying to remember without looking, testing yourself before you feel ready, reviewing material that has already started to slip away, consistently produce better long-term retention because of that difficulty, not in spite of it.

Bjork, R.A. (1994). Memory and metamemory considerations in training. In Metacognition: Knowing about knowing. MIT Press.
Study Design & Methodology PRE-REGISTERED STUDY Analysis plan was fixed before data collection began. Results reported as observed.

What this study is.

This section describes the study design, what it is testing, and what it is not. It was written before data collection began, and the analysis plan was fixed at that point. Results are reported as they are, not selected for the ones that look best.

Research Type Controlled study in real classrooms
Setting Pilot schools, India, replicable globally
Participants Class 6–8 · pilot n=5 · RCT n=90 (30 per arm)
Intervention Active recall + spaced practice repetition
Comparison Traditional passive revision
Data Source Per-attempt logging · live to your Passport
Abstract
What this project is trying to find out

Classrooms everywhere face a structural learning problem that predates technology, curriculum, and teacher quality: the way students review material actively works against how human memory consolidates. Standard revision, re-reading notes, oral repetition, massed practice before tests, is near-universal in classrooms of every kind. According to five decades of cognitive research, it is also among the least effective methods available.

Cognivia is a student-led research initiative that attempts to close this gap. It builds on two of the most robustly replicated findings in cognitive science, spaced repetition and active recall, and translates them into a lightweight tool that runs on a single shared classroom computer. The tool logs every student response, timestamps every session, and captures the data needed to measure whether the intervention actually changes how much students retain a week after instruction.

The core question is direct: does a structured, algorithmic review system produce measurably better retention one week after instruction than the traditional method it replaces? The study is ongoing. This page reports the methodology, the live data tool, and, as evidence becomes available, the results, including any findings that do not support the hypothesis.

Methodology
How the study is designed
Research Question Does error-taxonomy-aware adaptive learning, adapting each review to the type of error a student makes, produce greater 2-week delayed retention than correctness-only adaptive learning and standard classroom instruction, in Class 6–8 students in rural government schools?
Design Three-arm, parallel-group, pre-registered randomised pilot trial · randomisation stratified by school and class level (1:1:1) · pre-test / immediate post-test / 2-week delayed retention test
Arms · n = 90 (30 per arm) A, error-taxonomy-aware Cognivia with error-type-matched Hindi remediation · B, correctness-only Cognivia (adaptive by correctness alone) · C, standard NCERT classroom instruction, no platform
Participants 90 students · Classes 6–8 (ages 11–14) · three government junior high schools, Kanpur Dehat, Uttar Pradesh · randomly assigned after the pre-test
Measurement Points Week 0 pre-test · Week 6 immediate post-test · Week 8 delayed retention test, 2 weeks after post-test (primary)
Primary Outcome & Analysis 2-week delayed retention score · one-way ANCOVA with pre-test as covariate · three planned contrasts (A vs C, A vs B, B vs C), Bonferroni α = .025 · effect sizes (partial η², Cohen's d) reported regardless of significance
Secondary Outcomes Confusion Fingerprint Index (CFI-CF / CFI-INT / CFI-RF) predictive validity · immediate learning gain · Brier score (metacognitive calibration)
Limitations & Status Pilot trial powered for effect-size estimation, not confirmatory testing (n = 30 per arm; ~0.76 power for d = 0.40 on the primary A-vs-C contrast). Single investigator delivering the intervention. Pre-registered on OSF before data collection; analysis plan locked in advance. Independent ethics review, with verifiable written parental consent and student assent for every child; no response data is collected before that approval is in place. Reported per CONSORT 2010 with the pilot/feasibility extension. Results reported in full, null findings included with equal weight.
One Instrument, Every Learner

Designed for any learner.
Grounded in cognitive science.

Cognivia is not a classroom tool. It is an instrument for how memory works, made for anyone who needs to learn something and hold onto it. If you are thirteen or older, it was built for you.

Built for anyone who learns

Memory doesn't discriminate by age, curriculum, or institution. The same cognitive machinery, the same forgetting curves, the same retrieval benefits, operates in a schoolchild reviewing science and a postgraduate preparing for a high-stakes exam.

Cognivia was designed with resource constraints in mind, nothing to install, no setup, a single shared computer. But those constraints turned out to be advantages. A system that works for the hardest case works for every case.

School students, university undergraduates, self-directed learners, competitive exam aspirants, researchers, the underlying mechanism is the same. Cognivia adapts to the learner, not the other way around.

School students
Board preparation, subject revision, competitive entrance tests, built for students with limited time and no private tutors.
Self-directed learners
Structured, scientifically scheduled revision outside any formal institution, spaced practice without a classroom.
Undergraduates & graduates
Semester-long retention of dense technical content, adaptive intervals ensure material doesn't decay before exams.
Competitive aspirants
High-volume factual recall under time pressure. UPSC, JEE, NEET, CAT, Cognivia was built, in part, for exactly this.
How Cognivia works
First
Enrol, without a password

Enrol with a first name, age, subject, and an email. We send a one-time sign-in link, so there is no password to set. After that you return with your Cognivia ID. (The demo above needs none of this.)

Passwordless sign-in
Then
Get Unique Learning ID

A unique identifier is generated for every student. It links your session data, genome profile, and review history across all study sessions.

Persistent across sessions
Next
Choose your class and subject

Pick your class level and subject. Questions are drawn from Cognivia’s reviewed bank, every item tagged by topic and Bloom level, so what you are asked matches where you actually are.

Reviewed question bank
During
Cognivia Generates Adaptive Questions

Questions are drawn in order of memory urgency, what you are most likely to forget is reviewed first. The schedule updates after every response using the SM-2 algorithm with personalised ease factors.

SM-2 spaced repetition
Underneath
Learn Through Cognitive Science

Each question presents four options. Selection and response time are both logged. Distractors are engineered as adjacent concepts, so a wrong pick reveals which misconception is active, not just that one exists. Every mechanism maps to a published cognitive science finding.

Testing effect · Spacing · Error diagnostics
After
Your Passport updates

Every attempt is recorded: timestamp, response time, the confidence you rated, and the error type behind a wrong answer. Those feed your seven Learning Genome variables, and your Cognitive Passport is recomputed after every session.

Per-attempt logging
The System

Built around how memory works.
Not how teaching usually happens.

Runs on any shared classroom computer. Nothing to install, nothing to configure. Every feature maps directly to a mechanism in the cognitive science literature, nothing else.

Four-option recall with diagnostic distractors

Question → four options → selection + response time → feedback. Distractors are engineered as adjacent concepts, so a wrong pick reveals which misconception is active. Latency separates fast recall from cognitive load.

Cognitive principle: testing effect + error-type diagnosis

Review timed to the forgetting curve

Quality ratings 1–5, not binary. Personalised ease factor and review interval per question per student. Faithfully implements Wozniak (1990). No approximations.

Cognitive principle: spacing effect

Confidence rated on every answer

Every response carries a 1–5 confidence rating alongside it. The gap between stated confidence and actual accuracy is what exposes a confident error, the kind that survives revision because the student never doubts it.

Cognitive principle: metacognitive calibration

Every attempt logged automatically

Every attempt logged: timestamp, response time, quality rating, ease factor, next review interval, similarity ratio. 16 columns. Streamed live to your Passport. No aggregation.

Empirical principle: trial-level longitudinal tracking

Live dashboard

Accuracy trends, retention curves, and quality distributions render from your live attempt log instantly, no upload, no waiting.

Cognitive principle: metacognitive monitoring

Runs anywhere, no installation

Browser-based. Nothing to install, nothing to configure. Designed to run on a shared classroom computer, a laptop, or a phone, anywhere.

Deployment principle: access is a prerequisite for any intervention
Walkthrough

What happens inside
a single session.

A brief guide for anyone running Cognivia for the first time, student, teacher, or reviewer.

COGNIVIA, LIVE SESSION
SESSION PROGRESS 0%
Quick Start, 60 seconds
  1. 01Create an account with your name, age, and email, we send a one-time sign-in link
  2. 02Pick your class level and subject from the catalog
  3. 03Choose one of four options for each card, from memory, don’t look it up
  4. 04Cognivia scores each answer on accuracy and response time
  5. 05Watch your Cognitive Passport update after each session
01
Creating your account
Sign up with a first name, age, and email. We send a one-time sign-in link, no password to remember. Cognivia generates a learning ID that links your responses across every session, so you can pick up exactly where you left off from any device.
02
Answering each card
A question appears with four options. Pick the one you believe is correct, from memory, not by looking it up. Cognivia scores each response on accuracy and how long you took, then schedules when the card reappears. The faster the right answer, the longer the next review interval.
03
What happens in the background
After each answer, SM-2 silently recalculates this card’s review interval and ease factor. Hard cards return tomorrow. Mastered cards return in weeks. The session ends after twelve cards, a fixed length held constant during the trial so every student gets the same dose; flexing it to each student's fatigue comes once the study is done.
04
Your Cognitive Passport updates
Seven cognitive variables, forgetting rate, retrieval strength, response latency, confidence gap, consolidation, fatigue index, pattern dominance, recalculate after each session and write to your Passport. Nothing to download, nothing to drag anywhere.
05
Reading your trajectory
Focus on the accuracy trend line, a rising slope across sessions means spaced retrieval is working. A rising trend with falling response times means the knowledge is becoming automatic, not just recalled.
Intelligence Layer

Most tools tell you your score.
Cognivia tells you how your mind is learning.

Behind every session is a cognitive engine tracking memory strength, decay rate, and thinking speed
not to grade students, but to understand them.

Cognitive Dashboard
Memory Strength
Mapped Over Time
Tracks memory strength, forgetting speed, and response latency across every session. Trends surface automatically, no interpretation needed. Visualized like a health graph: clean, continuous, and built for understanding at a glance.
Memory Strength Forgetting Rate λ Response Time Retention Trend
Learning Diagnostics
Not What Went Wrong
Why It Did
Detects behavioral patterns including recall failure, overconfidence, and session-end fatigue. The system identifies whether a mistake is a knowledge gap, a timing problem, or a confidence miscalibration, and responds accordingly, within the same session.
Recall Failure Overconfidence Fatigue Index FSI Confidence Gap CG
Student Report
A Cognitive Profile
Not Just a Score
Generates a downloadable PDF report for each student, showing strengths, weak areas, topic focus, and exactly what to revisit next, and when. Includes the exact return date for each topic, calculated by spaced repetition. No summary required, the report reads itself.
Strengths Map Weak Topics Next Study Plan Return Date
Architecture

What runs inside
every session.

Every layer both consumes and produces structured data. The loop closes at session end and reopens smarter at the next session. SM-2 is the foundation, every layer above it modulates, never replaces it.

L7
ResearchEngine Every session enters the research dataset, one row per attempt, feeding four pre-registered hypotheses that are re-tested daily.
L6
LearnerFeedbackEngine Plain-language diagnostic report answering 4 questions every student needs, every session.
L5
DifficultyEngine + Router Bloom-aware empirical difficulty. Memory-state-weighted question routing per session.
L4
AdaptationEngine 6 memory states. Quality override. Post-SM-2 interval modulation. Within-session requeue.
L3
CognitiveDiagnostic 4-type error classifier: RECALL_FAILURE · PARTIAL_KNOWLEDGE · CONFABULATION · FATIGUE.
L2
ForgettingModel Per-learner, per-topic λ calibration. RS weighted mean. CE variance. All update every session.
L1
LearningGenome 7 cognitive variables forming a behavioral fingerprint per learner. The base of all intelligence.
SM2Engine Wozniak (1990) spaced repetition. Preserved intact, every layer above modulates, never replaces it. FOUNDATION
An uncommon combination
I

Genome-Modulated Scheduling

Every adaptive system applies one algorithm to all learners. Cognivia applies a personal coefficient, derived from 6 live behavioral signals, on top of SM-2. A SLOW-forgetter and FAST-forgetter receive structurally different schedules even when they score identically.

Final interval = SM-2 × f(λ, RS, CE, FSI, streak, memory_state)
II

Intra-Session Cognitive Diagnosis

Duolingo, Khan Academy, BYJU's, all defer adaptation to the next session. Cognivia detects FALSE_CONFIDENCE during the session and acts: requeue, interval compression, quality override. The intervention window is minutes, not 24 hours.

CognitiveDiagnostic → within-session requeue → live memory state chip in UI
III

Auto-Generating Research Instrument

Cognivia doesn't produce a teacher dashboard as a secondary feature. It produces a pre-publication research dataset as a first-class output. Every classroom running Cognivia is simultaneously generating data toward peer-reviewed hypotheses, without any manual data entry by anyone.

H1: d ≥ 0.3 adapted vs SM-2 · H2: λ clusters into 3 archetypes (ANOVA) · Target: Computers & Education IF≈11
The 60-second version

Most study tools tell you what you got wrong. Cognivia pays attention to something harder: why you got it wrong, and when you're about to forget it. As you answer, it tells apart plain forgetting, half-knowing, and being confidently wrong, and it works out how fast each idea fades for you in particular. Then it brings each thing back right before you'd lose it, instead of on a fixed schedule that ignores how your memory actually works. Everything below is how that gets measured, in detail. You never need any of it to use the app.

The core idea
Most learning systems treat students as if they learn in the same way
at roughly the same pace, with the same patterns of forgetting.
In practice, this is rarely true.
Some students forget rapidly but recover quickly. Some retain information longer but struggle to retrieve it under pressure. Some answer correctly, but only after long hesitation. Others respond quickly, with misplaced confidence. These differences are usually invisible. Cognivia attempts to make them measurable.

Two students. Same class. Same content. Different memory profiles.

Student A, retains slowly
recalls confidently
Student B, fast forgetter
high retrieval speed
What each axis measures
λForgetting rate
SRetrieval strength
τResponse latency
ΔConfidence gap
ηConsolidation
φFatigue index
πPattern dominance
A Different Unit of Learning

Memory behavior, not marks, not completion.

Instead of measuring learning through grades or progress bars, Cognivia models each student through a set of evolving variables, updated after every session, every response, every hesitation.

Together, these variables form a dynamic profile: a Learning Genome, a continuously updated representation of how a student’s memory actually behaves. Not how fast they finish, or what they scored. How their memory consolidates, decays, and retrieves knowledge over time.

Learning Genome
λ RS RLI CG CE FSI PDS
Per-student behavioral fingerprint
Updated every session · 7 live variables
What This Changes

Most systems ask only one thing of a wrong answer, whether it was right. Everything follows from that binary: uniform review schedules, passive re-reading that still counts as revision, and no way to tell a student who simply forgot from one who never understood in the first place. Cognivia asks harder questions, how stable is this knowledge, how fast is it decaying, is the student retrieving it or guessing, and does their confidence line up with reality. Each review, each response, each hesitation becomes a signal, and the system adapts not to the content but to the structure of the learner’s memory itself.

The Learning Genome

Memory is individual.
This is how Cognivia tracks it.

Several things happen inside every Cognivia session, all running automatically in the background. Each one is grounded in research on how memory works. Together they form a complete loop: the system learns how each student remembers, schedules each review at the right moment, detects when a student’s understanding is wrong rather than just incomplete, and records everything for research. It runs in the browser on a single low-cost computer, and a full session takes under twenty minutes.

How to read the math below

These seven formulas and their cutoffs are Cognivia's own operationalisations, not results borrowed from the papers cited across this site. The decay form draws on Ebbinghaus and on Wixted & Carpenter (2007); the storage-versus-retrieval distinction on Bjork & Bjork (1992); the confidence and error-type signals on the metacognition literature. But the exact expressions, the smoothing coefficients (α = 0.75, α = 0.30), and every threshold (λ ≤ 0.7, CE > 0.85, FSI > 0.20, PDS > 0.6) are engineering defaults we chose, seeded from the n = 5 pilot and the literature. They are how the system computes today, not validated constants, and not claims any cited author has made. Calibrating these thresholds against a real student population is one of the things the trial is for.

01
λ
Forgetting Rate
R(t) = exp(−λ · t / S)

How fast a topic fades for you. A student at 0.5 remembers twice as long as one at 1.0, even if they score the same today, so the two need different review schedules. We estimate it from your own answers and sharpen it with every review.

02
RS
Retrieval Strength
EWM(quality, α=0.75)

How reliably you can pull a fact back right now. You can score well today but still be weak here if you crammed yesterday, so it separates real mastery from a recent glance.

03
RLI
Retrieval Latency Index
t_student / t_cohort_median(diff)

How long your recall takes, next to others on the same difficulty. Fast and correct means it is becoming automatic; slow and wrong points to a genuine gap, not just hesitation.

04
CG
Confidence Gap
stated confidence − accuracy (EMA)

The gap between how confident you are and how right you are, read from the 1–5 rating on each answer. Confident and repeatedly wrong is not ignorance, it is a misconception, and it needs correcting rather than more review.

05
CE
Consolidation Efficiency
1 / (1 + var(λ_history[−6:]))

How steadily your memory settles after the first few reviews. High means recall is consistent and your projections are reliable; low means you might score 90% one day and 50% the next on the same material.

06
FSI
Fatigue Susceptibility Index
EMA(acc_drop_session, α=0.30)

How much your accuracy drops from the first third of a session to the last. A high value means the session should be shorter, not that you are learning less.

07
PDS
Pattern Dominance Score
max(recall_n, partial_n, confab_n) / total_errors

How much of your errors are the same type. A high value means you are failing in one specific way, not at random, so the fix is targeted correction, not more repetition.

Recall failure
Memory trace absent. No encoding or complete decay.
Partial knowledge
Encoding exists but is incomplete or fragmented.
Confabulation
Active misconception. Confident and wrong, requires explicit correction, not repetition.
Fatigue
Knowledge exists. Execution degraded by cognitive load near session end.
“If learning is treated as uniform, teaching remains generic.

But if learning can be measured at the level of memory behavior, it becomes possible to identify fragile understanding before failure occurs , and begin to map how different students actually learn over time.”
Identify fragile understanding before failure occurs
Detect misconceptions through patterns, not just errors
Adapt review timing to the individual, not the average
Map how different students actually learn over time
Cognivia does not assume how learning works.
It attempts to observe it directly, one student at a time.
Six Memory States

The genome feeds a six-state memory classifier. State determines how the SM-2 interval is overridden, not replaced. The base algorithm remains intact; the genome modulates it.

State Trigger Action on SM-2 Why
False confidence Fast wrong + high CG. Active misconception. Quality → 2  ·  interval × 0.5  ·  requeue now Most dangerous state. Must be destabilised before next review.
Strong Low λ + high RS + stable CE + fast retrieval. No override  ·  interval × 1.2–1.4 Genuine durable encoding. Extend confidently. Don’t waste review time.
Fragile High λ OR unstable CE OR correct-but-slow. No override  ·  interval × 0.7–0.85 Performance may be misleading. Compress interval to prevent false confidence.
Slow retrieval Correct but RLI > 1.8. Cognitive load, not mastery. Quality cap 4  ·  interval × 0.85 Slow correct = effortful recall, not yet automatic. Don’t over-extend.
Fast retrieval Correct AND RLI < 0.6. Automaticity confirmed. No override  ·  interval × 1.3–1.5 Deep encoding confirmed. Extend beyond SM-2. Redirect effort to harder material.
Normal No special signal. Standard λ/RS/CE variance. SM-2 × f(λ RS, CE) only No extraordinary signal. Genome modulation only.
Predict, then verify

Cognivia predicts whether you’ll remember.
Then it checks itself.

After each session the engine forecasts how likely you are to recall every concept on a future date. When that date arrives, it compares the forecast to what you actually remembered and records its own error. Most tools grade the student. This one is graded too.

Every forecast, logged
A prediction is stored for each concept before its recall is ever tested.
Checked against reality
When the review comes due, the forecast meets actual recall and the gap is measured.
Accuracy, when earned
The engine reports its own accuracy only after enough predictions are verified.
How the self-check is scored Illustrative example

The metric is the Brier score, the squared gap between a forecast and what actually happened, where 0 is a perfect prediction and 0.25 is a coin toss. If the engine says you have an 80% chance of recalling a concept and you do, that forecast scores 0.04; if it says 80% and you forget, it scores 0.64. A session's score is the average across all its forecasts. The pre-registered bar is simple: beat a naive baseline that just predicts the class's average recall rate. If a per-student model can't beat that, it isn't earning its complexity. The live number stays blank on purpose until enough forecasts have been verified, this is the one figure the page won't show before it's real.

Cognitive Passport

A mathematical picture
of how your mind learns.

After enough sessions, Cognivia knows things about your memory that you don't. Your optimal study time. Which subjects you forget fastest. Whether your confidence tracks your actual knowledge. The Cognitive Passport surfaces all of it, in one page, updated after every session.

Cognitive Passport
A grade tells you what you got right.
The Cognitive Passport tells you who you are as a learner.

Every score you've ever received collapses an enormously complex cognitive process, weeks of encoding, forgetting, retrieval, consolidation, into a single number. The Passport doesn't. It captures the full geometry of how you learn, so that for the first time, a document can exist that reflects not your performance on one day, but the shape of your mind across all of them.

COGNIVIA COGNITIVE PASSPORT ISSUED BY COGNIVIA LEARNING SYSTEMS · 2026
DOCUMENT NO. CGV-XXXX-SAMPLE
Student Profile
Aarav M.
λ RS CG φ τ η PDS
Study Duration 47 days
Cards Mastered 183
Accuracy Trend ↑ +12%
Cognitive State CONSOLIDATING
System Recommendations
Extend Chemistry intervals, RS above 0.85 threshold
Peak encoding window 10:00–12:00, schedule your hardest topics here
Confidence calibration needed, CG gap widening in Physics
Sample · Generated in-browser · No data stored

The Cognitive Passport is recomputed after every session, reflecting changes in your Genome variables as they happen. Unlike any transcript or marksheet, it captures how you learn, not just whether you passed.

A transcript records performance on a single day, in a single room, under artificial pressure, then averages it into a number that follows you for life, none of which reflects how you think. The Passport records what that number hides: how quickly you forget, how fast you retrieve, how closely your confidence tracks your accuracy, the patterns in how you consolidate, even the times of day you encode best, the shape of how your memory works.

Submitted to a university, a document like that could tell admissions something a GPA never can, that a student doesn't just know the material but understands how their own mind works. And where coaching and tutors are unequally distributed, it levels the field: it is evidence of learning rather than purchased preparation.

How Your Passport Is Generated
01
You studyAnswer cards, rate recall quality
02
Every response loggedTimestamp, latency, confidence, accuracy
03
Genome updatesGenome variables recalculated locally
04
Passport generatesPDF rendered in-browser, on your machine
Controlled Study

A controlled study.
Built to be replicated.

Cognivia is not making promises. It is being tested under the specific conditions it was built for, Class 6 to 8 students in a low-resource setting, studying on a shared classroom computer, with no prior exposure to structured review methods. Those conditions are very different from a university laboratory, and whether the science actually holds in a setting like this is exactly what we are trying to find out.

The Central Question

“Does error-taxonomy-aware adaptive learning, adapting each review to the type of error a student makes, produce greater 2-week delayed retention than correctness-only adaptive learning and standard classroom instruction, in Class 6–8 students in rural government schools?”

Independent Variable
Adaptation signal across three randomly assigned arms: error-taxonomy-aware Cognivia (A) vs. correctness-only Cognivia (B) vs. standard NCERT classroom instruction (C). Arms A and B are matched on everything except what the system does with a wrong answer.
Dependent Variable
Score on parallel-form 30-item NCERT Science assessments at Week 6 (immediate post-test) and Week 8 (2-week delayed retention test), the pre-registered primary outcome.
PhaseWeekStandard instruction (C) Cognivia arms (A & B)Measurement
Baseline Week 0 30-item NCERT pre-test Same 30-item pre-test Prior knowledge score
Instruction Weeks 1–6 Standard NCERT teaching 18 Cognivia sessions (3×/week, 45 min) Per-attempt logs streamed to Passport
Post-Test Week 6 Same instrument Same instrument Immediate learning gain
Delayed Test Week 8 Same test; no review since post-test Same test; no review since post-test Primary: 2-week delayed retention
H₁ · Primary

Memory After Two Weeks

“Group A (error-taxonomy-aware) will show higher 2-week delayed retention than Group C (standard instruction) and Group B (correctness-only), controlling for pre-test (one-way ANCOVA; planned contrasts, Bonferroni α = .025; Cohen’s d).”

H₂ · Secondary

The Confusion Fingerprint

“The Confusion Fingerprint Index (CFI-CF), computed from early sessions, will predict 2-week retention above and beyond overall accuracy, validating it as an individual-difference construct (hierarchical regression; F-change).”

Group Comparison

How the arms compare.

The trial randomly assigns students to three arms. Two use Cognivia, one adapting to the type of error a student makes, one adapting to correctness alone; the third is standard NCERT classroom instruction. All three cover the same content, in the same schools, over the same six weeks. The charts below show what the data records.

Group C · Control
Standard Instruction
Students taught the standard NCERT way: teacher-led instruction, re-reading, oral repetition, and written exercises. No adaptive tool, no individual tracking, no platform. Same subject content, same schools, same six weeks as the Cognivia arms.
30 Students
Standard NCERT How they study
None Records kept
Groups A & B · Cognivia
Adaptive With Cognivia
Two randomly assigned Cognivia arms, three sessions a week, 45 minutes each, on a shared classroom computer. Students answer from memory, rate their confidence, and move on; Cognivia logs every response and reschedules each question. The two arms differ in one thing: Group A adapts to the type of error (error-taxonomy-aware, with matched Hindi remediation); Group B adapts to correctness alone. That contrast isolates the mechanism.
60 Students (30 + 30)
Adaptive review How they study
Full log Records kept
What the data showed
The comparison is fixed in advance: 2-week delayed retention across the three arms, analysed by one-way ANCOVA with pre-test as covariate and three planned contrasts (A vs C, A vs B, B vs C; Bonferroni α = .025), reporting effect sizes for every contrast whether or not they reach significance.
Why this result makes sense
The established research on memory suggests the error-taxonomy arm should retain most. When students practice retrieving information rather than re-reading it, and when review is matched to the specific way a wrong answer went wrong, the memory trace formed is stronger and more resistant to forgetting. That is why Groups A and B are matched on everything except what the system does with a wrong answer: the A-vs-B contrast isolates the mechanism from the general benefit of adaptive practice. This has been shown in laboratory studies. What we do not know is whether it survives the translation into a rural Indian classroom, with shared devices, irregular attendance, and an exam-driven syllabus that leaves little room for experimental methods.

If the study finds no meaningful difference between the arms, that is not a failure but equally important information: it would tell us that something about this context prevents the effect from appearing, and that the approach needs adjustment before scaling.
Roadmap

What happens next, and when.

The trial has a closing date. If the scheduling result comes back null, that one claim is reported as null and set aside; the data stays public, and the measurement layer beneath it, the Genome and the Passport, carries on as a diagnostic in its own right.

Feb – Apr 2026
India pilot · n = 5 · closed
One rural school in Kanpur. 548 questions answered, 36 sessions logged. Feasibility study only, not a result.
Jun – Aug 2026
Three-arm RCT recruitment · n = 90 (30 per arm)
1:1:1 allocation, stratified by school and class level, concealed until after the pre-test. Pre-registered exclusion criteria. Recruiting through government schools in Kanpur Dehat.
Sep 2026
2-week delayed retention collected (primary)
Primary outcome data collected for every enrolled participant. No further intervention after this date, only the probe.
Oct 2026
Pre-registered analysis runs · OSF posted
The analysis script (already public in the repository) runs once on the locked dataset. Results posted to OSF regardless of outcome, positive, null, or negative.
Nov 2026
Journal submission
Primary target: Computers & Education or Journal of Educational Psychology. Submission goes out regardless of result; a null finding is a publishable finding.
Dec 2026
The mechanism bet resolves
If the effect appears (Cohen's d ≥ 0.3, pre-registered), planning begins for an independent replication run by someone else, somewhere else. If it does not, the scheduling claim is reported as a null and set aside, while the Genome and Passport measurement layer carries on. The repository, the data, and the protocol stay public, that is the only honest version of research.
Public record

Preprints & pre-registrations.

Every claim here traces to a public, citable record, pre-registered before data, with preprints and code openly available.

Retention study · pre-registration
OSF · 10.17605/OSF.IO/YU6P5
Retention study · preprint
Zenodo · 10.5281/zenodo.20680880
Retention study (v1) · preprint
EdArXiv · 10.35542/osf.io/cr53e_v1
Confusion Mapper · preprint
EdArXiv · 10.35542/osf.io/5k9rv_v1
Confusion Mapper · archive
Zenodo · 10.5281/zenodo.20807432
Also indexed on SSRN
SSRN 6933798  ·  SSRN 6933638
Code & data
github.com/Manik-Maurya
Pre-registered, tracked in public

Three predictions that count.
One worth watching.

Four hypotheses were registered on OSF before the trial began. H1–H3 are confirmatory, evaluated with Bonferroni-corrected planned contrasts (α = .025). H4 is explicitly exploratory, reported with effect sizes but never used to confirm or disconfirm the primary conclusions. Keeping the confirmatory set small is deliberate: piling on corrected tests at this sample size would be a formality, not an analysis.

Power, stated plainly. This is a pilot trial powered for effect-size estimation, not confirmatory certainty. With n = 30 per arm, the primary contrast (A vs C) has about 76% power to detect the conservatively assumed effect (d = 0.40) under a Bonferroni-corrected ANCOVA. A null result here narrows the plausible effect; it does not rule out a smaller one, and this page will not claim otherwise. The adequately powered confirmatory trial is a later, larger study.

A = error-taxonomy-aware Cognivia · B = correctness-only Cognivia · C = standard instruction
H1
Confirmatory
Does adapting to the type of error a student makes produce more 2-week retention than correctness-only adaptation and standard teaching? Awaiting data
H2
Confirmatory
Does a student’s Confusion Fingerprint predict what they retain, beyond their overall accuracy? Awaiting data
H3
Confirmatory
Does error-aware feedback improve calibration, closing the gap between how sure a student feels and how right they are? Awaiting data
H4
Exploratory
Is the Confusion Fingerprint a stable trait, and does error-taxonomy remediation change it? Awaiting data
Syncing with the live data…
548
questions answered
36
sessions logged
5
students (feasibility)
7
subjects · India pilot · 2026
Feasibility pilot · n = 5

What the pilot showed, honestly.

These are the real results from the pilot study, 548 questions answered, 36 sessions, 5 students across 7 subjects, from February to April 2026 in India. They were collected on the earlier desktop prototype, not on this web platform, and are reported here as a feasibility check rather than as trial evidence. Every number was computed directly from that pilot log. If something did not work, it is reported here the same way as something that did.

Pilot result · Feb–Apr 2026 · India
+44.8pp
Accuracy advantage · Cognivia vs standard revision
Pilot · n = 5 · feasibility study only
Cognivia group
81.7%
Standard revision
36.9%
Projected · multi-school RCT
+18.0pp
at n = 90
With n = 5, this is a descriptive difference only, too few for a meaningful confidence interval or effect size. Those are reported on the full pre-registered sample, not the pilot.
Pilot effects from small samples shrink toward the truth as n grows. Adjusted for regression-to-the-mean and lab-to-classroom decay. Pre-registered analysis runs on the locked dataset in October.
Informal pilot  ·  Feb–Apr 2026
Data from the full pilot: 548 answered questions, 36 sessions, 5 students, 7 subjects. India. February to April 2026. All numbers computed directly from the raw session logs.
February – April 2026  ·  548 questions answered  ·  5 students total  ·  3 using Cognivia, 2 using traditional revision  ·  Results tested for statistical significance
Before & After

What changed.
Within the same students.

This section compares the same students against themselves, their performance in the first half of the study against their performance in the second half. Both groups are the same people using the same tool. The only thing that changed is time, and how many times they had reviewed each piece of material. If the numbers improve, it is because practice and spaced repetition actually worked.

Pilot in progress  ·  First cohort is mid-way through the protocol. Real numbers will replace this preview once the cohort's results are uploaded.
Average accuracy, before
First 3 sessions, group mean
Average accuracy, after
Last 3 sessions, group mean
Group-level change
Percentage-point shift
Students who improved
Share of individually-tracked learners
Retention (3+ reviews)
Accuracy on cards seen 3+ times
Students with sufficient data
At least 8 logged sessions
Before vs After accuracy
Group mean by phase
Session trajectory
Session-by-session with phase boundary
Gain distribution
Count of students by improvement bracket
Once the protocol completes, this panel will populate automatically with the cohort’s actual results, including any finding that contradicts the hypothesis. Null results are reported with the same weight as positive ones.
Limitations

What this study doesn’t claim.

Every study has limits. Stating them clearly is not weakness; it is the only way the results carry weight. The limitations below are not buried in footnotes. They are the conditions under which anything in this study can be trusted.

What may not work

What works in a laboratory may not work in a real classroom

Decades of research have shown that spaced practice and retrieval testing improve memory. Almost all of that research was done with university students sitting at individual computers in quiet rooms. Whether the same effects appear in a resource-limited school, with many students sharing one device, irregular attendance, and teachers under pressure to cover a fixed syllabus, is an open question. We do not assume the answer.

There may be subjects, age groups, or classroom conditions where these methods simply do not work as well. A result that shows no difference between the two groups would be reported here with the same weight as a positive finding.

What is still uncertain

Five students is not enough to draw broad conclusions

This pilot involved five students from one school. That is too small a group to make any general claim about how students learn everywhere, or even in all classrooms like it. What it can do is establish whether the method is worth testing more rigorously, with a larger, properly randomized group spanning different countries and contexts.

The students who completed fewer sessions have less data behind their individual profiles. Those profiles are less reliable and should not be used to make strong predictions about how those specific students will perform.

What needs further validation

The individual student profiles are theoretical at this stage

Cognivia tracks seven variables about each student’s learning behaviour and builds a profile from them. Whether that profile can reliably predict who will struggle or identify the best time to intervene has not yet been proven. This study begins to generate the kind of data needed to test that.

Response speed is recorded and used as a rough measure of how fluently a student can recall information. But on a shared device, speed is also affected by how fast the computer responds on a given day. This is a real confound that cannot be fully eliminated in this setting, and it is factored into how we interpret those numbers.

A confound we can't design away

The teacher knows which class is piloting the new tool

Both groups share the same teacher, subject, and week, which controls for a lot. What it does not control for is enthusiasm: a teacher who knows one class is using new technology may, without meaning to, teach it with more energy or attention. That expectancy effect would inflate the result independently of the tool. Full blinding is impossible here, so we treat it as a live threat to validity, plan a teacher-behaviour check, and read the two-arm technology comparison (A versus B, where both classes use Cognivia) as the cleaner test precisely because it holds that enthusiasm roughly constant.

What happens when attendance is irregular

Students will miss sessions, and some will drop out

In the classrooms we work in, irregular attendance is a fact, not an edge case, so the analysis is pre-specified for it rather than surprised by it. The primary analysis is intention-to-treat: a student is analysed in the arm they were assigned to, whether or not they completed every session. The number of sessions attended is recorded as a dosage covariate. Missing post-test or retention scores are handled by a pre-registered rule (multiple imputation under a missing-at-random assumption, with a complete-case sensitivity analysis reported alongside), and attrition is reported per arm in a CONSORT flow diagram. If dropout is heavy or unbalanced between arms, that is reported as a limitation on the result, not smoothed over.

Forward Path

What becomes possible
if the results hold.

None of what follows is certain. Each possibility depends on the results being real, the methodology holding up under review, and the work being replicated at a larger scale. Each depends on the evidence holding up at a larger scale.

Any teacher anywhere in the world could replicate this study

The study design, the data format, and the analysis approach are all documented and publicly available. A teacher in a different city, working with different students and a different subject, could run the same experiment and compare the results. The value of this kind of work multiplies with each independent replication.

A rare kind of data, from classrooms that are rarely studied

There are very few long-term, question-level datasets tracking how real students in low-resource schools actually forget and retain material over time. If this study scales, even modestly, it begins to fill a gap in the research that no single experiment can address alone.

Submitting the findings to peer review

The methodology, results, and limitations are being prepared for submission to a peer-reviewed journal. A negative result will be submitted alongside a positive one. The credibility of the work depends on being willing to publish both.

Giving teachers information they can actually use

The dashboard already shows which students are struggling with which subjects, and how their performance is changing week by week. Developing that into something a teacher can consult at the start of each lesson, without any technical knowledge, is a natural next step.

More classrooms, only if it actually works

If the results hold up and replication supports them, Cognivia can be deployed in additional classrooms without any additional infrastructure. It runs on a single computer, with nothing to install or configure. The conditions that make it practical in a resource-limited school are the same ones that make it scalable, globally.

The data matters more than the tool

Cognivia is one tool. The more durable outcome is the evidence it collects about how students in underfunded schools actually learn and forget, evidence that does not currently exist in the research literature. That is what has the potential to influence how teaching is approached at a systemic level, in ways no single application ever could.

Where this
comes from.

Independent research  ·  2022–2026

Across classrooms around the world, the same pattern recurs: students work hard, revise consistently, and still forget most of what they reviewed a week later. The problem was never effort or ability, it was the method.

Nothing about the way these students were being asked to revise was aligned with how memory actually works. The science of spacing, retrieval, and forgetting curves has existed for over a hundred years. It had simply never reached the classrooms where the need was most acute.

So I built Cognivia. It is not a quiz app or a learning platform; it is an instrument that models how each brain forgets and steps in at precisely the right moment, whether that person is in a rural school in India, a library in Lagos, or a bedroom in São Paulo. It runs on a single shared device, with nothing to install, because the classrooms that need it most cannot depend on heavy infrastructure.

A detour worth admitting: the first version was wrong. I spent months building teacher dashboards before I had anything worth putting on them, I had confused a product that looked serious with one that measured something. The measurement had to come first. Most of what Cognivia is now came out of scrapping that early build and starting from the memory science instead of the interface.

I locked the study design before collecting any data, and I report the analysis as it comes, including the results that contradict what I predicted. A null finding gets published with the same weight as a positive one. I would rather know the truth than protect the idea.

I did not start this to launch a product. I started it because a question would not leave me alone, and I built it where that question is hardest to answer. If it works, there is plenty to build on top of it, but the science has to come first, or none of the rest is worth trusting.

,  Manik Maurya
INDEPENDENT RESEARCH  ·  2022–2026
Evidence Base

The protocol is fixed. The result is not.

The shortest version of what Cognivia is for: The world teaches roughly 1.5 billion students in school every year. A tiny fraction of them are reached by anything resembling adaptive instruction. The classroom is not going to be rebuilt around each individual learner, not in any country, not in any decade I will be alive for. What can be rebuilt is the few minutes of revision each of those learners does on their own.

So this study asks a narrow question, does adapting the schedule to the type of error a learner makes improve retention? because if the answer is yes, the answer is portable. It runs on a phone or a shared classroom computer, and needs no teacher to install it and no school to approve it. It is, in the most literal sense, the only intervention I know how to build alone, from Kanpur, that could matter to a learner I will never meet.

This trial tests one falsifiable claim, that adapting to the type of error beats plain spacing. I report the result either way, with the data and protocol public. A null would mean that specific claim was wrong, not that the work ends, because the measurement layer beneath the scheduling, the Genome and the Passport, is useful as a diagnostic on its own, whatever the scheduling result.

The bigger picture, in plain terms

Honest about the harder questions.

What's open, and what isn't.

The protocol, the analysis, and this trial's data are public, that is how the science earns trust and how the mission stays honest. What is not open is the part that compounds: the trained models. Every learner sharpens the per-topic forgetting priors, the distractor banks tuned to real confusion patterns, and the cold-start models that let a new student be understood in a handful of questions. Anyone can read the methodology; the calibrated dataset and the models built on it improve the more the system is used. Open science and something worth building are not in tension here, they are separated on purpose.

Who owns the work.

To settle it plainly: the code, the trained models, and the calibrated dataset are independently owned. They are not the output of any institution, lab, or employer, and no one else holds a claim on them. The research is run independently and kept separate from anything built on top of it, so there is no tangled ownership to unwind later. Clean from the start.

Where Cognivia sits.

Each of these is stronger than Cognivia somewhere, in distribution, polish, or content. What none of them do is diagnose the type of error and schedule against a per-student forgetting model on low-end hardware, which is the position Cognivia is built to own.

How the content actually scales.

The engine is subject-agnostic; the bottleneck is the question bank. Every subject needs items authored, and every wrong option tagged to the specific misconception it stands for, because that tagging is what lets the system read error type instead of just right-or-wrong. It is slow, expert work. Today the bank is CBSE maths and science for Classes 6–8, written against the NCERT sequence and checked by practising teachers before anything goes live. Scaling is a pipeline, not a prompt: draft from the curriculum, tag the misconceptions, teacher review, pilot, release. And it has to be rebuilt per board, since CBSE, ICSE, state boards, and international curricula order topics differently. I would rather own one board deeply than claim “any subject, anywhere” I cannot yet stand behind. As the catalogue widens past what I can author alone, each domain gets its own subject-matter reviewer, a chemistry item and a history item need different expertise, and nothing reaches students without that domain sign-off. Content quality cannot rest on one generalist once the subject list is long.

Whether the questions actually measure what they claim.

A difficulty label is a hypothesis until the data backs it up. Right now item difficulty is assigned from Bloom level and expert judgement, and each distractor is designed to catch a specific misconception, but neither claim has been validated against real response patterns yet, and I won't pretend the difficulty engine has proven itself. The honest test is empirical: does a question tagged “hard” actually get missed more often, and does each wrong option get chosen by the students who hold the misconception it was built for, rather than at random? That is item-difficulty calibration (item-response theory) and distractor-functioning analysis, and it needs real answers at scale to run. The trial is what produces them. Until then the difficulty engine is a well-motivated design, not a validated instrument, and a page this quantified everywhere else should say so here.

How a teacher runs a whole class.

A twelve-year-old on a shared classroom computer should not be self-registering. The flow is meant to be teacher-first: the teacher creates the class, adds the roster in one action, assigns a subject to everyone at once, and gets one class view of who is struggling and on which error type, not thirty separate sign-ups. I will be straight about the gap: today’s live path is still per-student registration, and the teacher-held roster with a class dashboard is the next thing I am building, not something already shipped. The school case only works when the teacher holds the roster, so that is where the product goes next.

The credential problem, named.

If a Passport is ever going to sit next to a GPA, the hard part is not the model, it is trust: stopping a friend from doing someone's sessions, or gaming the confidence ratings. That means identity verification, anomaly detection on response patterns, and an audit trail, the standard infrastructure any credential needs. It is not built yet, and I will not pretend it is. It is the gate between a useful diagnostic and an accepted credential, and it sits on the roadmap deliberately.

And even solved, trust is only the technical half. Credentialing is a gatekept, institution-controlled space: boards, universities, and accreditation bodies do not adopt a new format because it is good, they adopt it slowly, through policy, pilots, and politics. Treating “replace the transcript” as a product-design problem alone is the naive version. The honest path is narrow and long, start as a supplement a student volunteers alongside a transcript, earn predictive-validity evidence a few institutions will actually look at, and let adoption follow proof rather than ambition. I would rather say that plainly than pretend an incumbent credential falls to a better UI.

Who owns the Passport, and what if Cognivia disappears.

A document meant to span years has to outlive whoever built it. So the rule is simple: the learner and their family own their Passport, not us. You can export the full record at any time in an open, machine-readable format, not a screenshot, the underlying data. And I have said elsewhere on this page that if the trial comes back null the scheduling claim is set aside; the same honesty applies to the data. If Cognivia ever changes hands, pivots, or winds down, personal records are returned to families and then deleted, the anonymised research set stays open, and a family's years of data are never sold off. A credential you cannot take with you is not a credential; it is a hostage.

What Cognivia will not do.

A project that leans this hard on scientific integrity should be just as disciplined about how it holds attention. So, plainly: no guilt-trip streaks, no loss-aversion notifications engineered to make a child anxious about a broken chain, no manufactured urgency, no dark patterns tuned to maximise time-on-app. Those tactics lift engagement metrics and corrode the thing we claim to serve. Cognivia should be able to be put down. What replaces the tricks is the actual mechanism: a reminder arrives when the memory model says a review is due, because that is when it helps, not because a growth loop needs a daily open. The measure of success is what a student remembers weeks later, not how many minutes they spent in the app today.

The first two weeks, before the science kicks in.

Here is the real retention problem, named honestly: the deep signal, a stable forgetting curve, a confident read on how a mind learns, only emerges after dozens of sessions. So a twelve-year-old on day three has earned almost nothing from the part that makes Cognivia special, and “come back in a month and it gets good” is how you lose them. The answer is not a streak or a badge; it is to make each early session pay off now, without faking it. Every session ends with something true and immediate: the concepts you locked in today, the ones scheduled to come back and when, and one thing the system noticed about how you answered. From the very first session the Passport isn't blank, it fills in live, one variable at a time, so the student watches their own portrait being drawn rather than staring at empty gauges. And the only reason to return tomorrow is an honest one: these specific cards are due, because that is when recalling them actually works. Early wins that are real, not manufactured, that is the bridge across the first two weeks.

What I would build the team around next.

I have built this alone so far. The first hires are the gaps I can already name: machine-learning and data engineering to run the calibration at scale, someone who has sold into schools and coaching institutes, and curriculum design to extend past the current subjects. Advisors in learning science and edtech distribution come before any of that.

The bigger bet.

The largest version of Cognivia is probably not one more learning app. It is the measurement layer other products call: an API any quiz app, tutor, or LMS can use to get error-type classification and forgetting-curve calibration for its own users, the way Stripe became infrastructure for payments rather than another store. That position owns the cross-platform data and is the hardest thing to copy. It is years out, and it is a direction, not a promise.

Proven in India, designed to generalize, and honest about the gap.

The India story is the strongest, most real thing here, and it stays front and centre, Kanpur, government schools, NEET and JEE and CBSE, a founder from exactly the context the product serves. But a visitor from Nairobi or São Paulo should read this and know the truth: it is proven in India and built to generalize, not already global. The universal part is the cognitive science, forgetting curves and retrieval practice work in every language. Almost everything else is real, hard, multi-year work I would rather name than gloss:

So the honest positioning is not “global today.” It is: prove the mechanism where I have the deepest context and the most authentic right to build, then generalise deliberately, one region’s content, calibration, compliance, and proof at a time. And the second market is a deliberate choice, not “everywhere”, the working plan names Nigeria (WAEC secondary), precisely because a different curriculum, a data-cost-sensitive network, and a different regulator (NDPA) force the platform to genuinely generalise rather than paper over the gap, while English-medium holds one variable steady. The full sequence is written down, not left to a hand-wave.

Straight answers

Questions people actually ask.

For students, parents, and teachers. If your question isn't here, write to us, the address is at the bottom.

Is it free?

Yes. Cognivia is free for students, and deliberately so for low-income learners, the whole point is to reach classrooms that can't pay. The research protocol, analysis, and trial data are open too. If a paid tier ever exists, it will be for institutions, never for a child trying to study.

How is my child's data used?

Only what the learning needs: which answers were right, how long they took, and the confidence rating on each one. Records are pseudonymised (a code, no names), stored on access-controlled servers, never sold, and never used for advertising. For anyone under 18, nothing is processed until a parent or guardian gives verifiable consent, as India's DPDP Act requires. Full detail is in the Privacy Policy.

What happens if we stop?

You can withdraw at any time, no reason needed. Your identifiable record is deleted within 30 days, and you can ask to see or remove your data whenever you like. There is no penalty and nothing follows the student afterward.

Will it work on a slow connection or a cheap phone?

That's the design target. Cognivia runs in the browser with nothing to install, the pages are light, and a session is a handful of small requests rather than video or large downloads, so it holds up on a basic phone and a weak connection. A full offline mode that caches a session and syncs later is on the roadmap; I'd rather build it properly than claim it before it's real.

What if the connection drops in the middle of a session?

Nothing is lost. Every answer is saved the instant it's submitted, not at the end, so if the connection drops on question seven, the first six are already stored. The session stays open, and when the student comes back it picks up where they left off instead of restarting. For the research, a partial session is still a valid, resumable session, so a flaky rural connection never silently costs a data point.

Can I put Cognivia on the home screen, like an app?

Yes. Cognivia is an installable web app: on Android or a laptop you'll see an “Add” prompt, and on an iPhone or iPad you add it from Share → “Add to Home Screen.” It then opens from an icon in full screen, no browser bar, no URL to remember, which matters a lot when the device is shared and the user is twelve. Nothing to download from a store, and it stays light.

What does it look like before I've done anything?

Not a wall of empty gauges. A brand-new account doesn't show a broken-looking dashboard of blank variables; it shows a clear first step, “Your passport opens after your first session”, and then fills in live as you answer, one variable at a time. The system is honest that it needs a handful of answers before it can say anything real about your memory, and it says so plainly instead of showing zeros pretending to be data.

We share one classroom computer, how do students switch?

This is the real problem in the classrooms we target, so it's designed around a teacher's roster: a student picks their name or a short code to begin their own short turn, and the previous session closes first so no one's answers ever log under someone else. The roster-based switch is part of the teacher flow I'm building next; today's single-student sign-in works, but the fast in-class handoff is not shipped yet, and I won't pretend it is.

Why does every session end at the same number of cards?

Fair catch, it looks inconsistent with a system that adapts to each student. During the trial the session length is held constant on purpose: to measure whether the method works, every student has to get the same "dose," or you can't compare them. Outside the trial, letting session length flex to each student's fatigue signal is exactly where it should go, and that's the plan once the fixed-dose study is done.

What languages is it in?

The interface is in English and Hindi today. Being honest about the harder part: the question content itself has to be authored per language, and many government schools teach in a regional medium, Tamil, Bengali, Telugu, Marathi, and more. Extending the actual item banks to those languages is a content-operations goal, not something already done. Translating the buttons is easy; translating a misconception-tagged question bank is the real work.

How does a teacher run this in a 40-minute period?

The short version: set up the class once, assign a subject, and let students take short turns while you watch a single class view of who is struggling and on which kind of error. A proper teacher guide, setup, a period-by-period routine, and change-management notes, is being written, because classroom tools usually fail on adoption long before they fail on technology. If you're a teacher and want to bring this to your class, write to me and I'll walk you through it directly.

Something's broken, or I have a question. How do I reach a human?

Email manikmaurya.in@gmail.com. It reaches me directly, and I reply within a few days. Teachers and parents are welcome to write before signing anyone up.

Start the story The Problem