Building Veritas, a learning app that checks what you remember

Veritas started because we forget much of what we read.   When I looked into why, the research was clear.   Most of what people do to study (rereading, highlighting, summarizing) barely helps.   The things that do help are testing yourself, spacing your practice and explaining things in your own words, and those are hard to keep up alone.

So I’m building one product that handles the dull parts.   You bring in a book, lecture notes or a long report.   Veritas reads it to you in your own voice, checks what you took in at the end of each chapter, and brings back what you missed a little before you’d have forgotten it.   I do the product work myself, which covers the research, the design of every step, the decisions about where AI belongs and the launch plan.

With the same passage and the same study time, people who read once and recalled it three times remembered 61% a week later. People who reread it four times remembered 40% (Roediger and Karpicke, 2006).

Three findings shaped everything that followed.

First, reading itself is in decline.   A 2025 study in iScience (Bone et al.) found that the share of Americans who read for pleasure on a given day fell from 28% in 2004 to 16% in 2023.   A 2026 report from Cornell’s Committee on the Future of the American University said students raised on digital technology are quick to find information but struggle with sustained reading and complex arguments.

Second, much of what people do with what they read doesn’t work.   In a 2013 review of ten common study techniques, Dunlosky and colleagues rated rereading, highlighting and summarizing as low utility.   Practice testing and spaced practice rated high.

Third, forgetting is predictable.   Ebbinghaus described the curve in 1885 and Murre and Dros replicated it in 2015.   Material you never revisit fades quickly, and each well-timed review flattens the curve.

Put together, the evidence pointed me away from faster reading and toward practice.   People need to rehearse remembering the things they have to know, at the right moments.

I started with three groups, since they read for different reasons and put up with different amounts of effort.   I’m building for the first one first.

I looked at four kinds of tools: read-aloud and audiobook apps, note-taking and study tools that work on your own files, flashcard apps, and quiz and game apps.   Each is good at its own step.   None of them knows what you missed in the step before.

The last three rows matter most.   A flashcard app can space review well, but only for cards you made.   A read-aloud app can read anything and has no idea whether you understood it.   I think there’s room in connecting the two, so what you get wrong in a chapter check shapes tomorrow’s review.

Before I drew a screen I wrote down the rules I’d use to say yes or no to features.   They come from the research.

The research points to three things for long-term memory: encode the material well, practice retrieving it, and space that practice out.   Veritas runs all three on anything you read.   Below is how each step works and what I decided along the way.

It starts in the library, which takes PDFs, EPUBs, Word files, web links, photos and scans, or pasted text, organized on shelves and in collections.   Chapters are found for you.   Then the reader hears the text with every word highlighted and a picture of the idea building beside it, following Mayer’s multimedia principle that people learn better from words and pictures than from words alone.

The speed test is a small decision with a big effect.   Most apps let you pick a playback speed and leave it, but people overestimate how fast they can listen and understand.   So the test reads short passages at increasing speeds, each followed by a comprehension question, and plots the result.   It shows where understanding starts to fall, then how long the current chapter and the whole book will take at that pace.   Chapter recaps are also read aloud in the reader’s voice, with sections they can jump between.

Every chapter ends with a knowledge check of about 12 mixed questions that read like a real exam, diagrams included.   It takes about 12 minutes.   After each answer the product explains why a wrong option is wrong, then why the right one is right.   Written questions ask the reader to explain something in their own words, and Veritas says what they covered and what they missed.

The weak points then drive practice.   The reader can work on them in six main ways: written answers, multiple choice, drawing from memory, teaching the topic back out loud, flashcards and games.   Veritas recommends one based on the reader’s own results, and the reader can reorder the list.   I included drawing and teaching back because the evidence for them is strong.   Explaining an idea, to yourself or to a friend, exposes gaps that recognizing the right answer hides.

The third step is meant to turn the product into a habit.   Instead of a review queue for every book there’s one daily review across all of them, in a calm full-screen mode.   What you miss comes back sooner and what you know comes back less often.   Before you start, the Today screen says what’s due, how many questions it is and about how long it will take.   I figure a review you can’t size up is one you put off.

Learners only ever use the web app in their browser: the library, voice setup, read-along, check and practice, and the daily review.   Sign-in is hosted, so Veritas never stores a password.   The browser also records the voice clips, and it asks for consent before it does.

Everything else runs in the cloud.   The web server checks who is signed in, shows each learner only their own data, and hands slow work to a job queue.   The queue is a table in Postgres, so jobs survive a restart.   Four workers take jobs from it: import (OCR and Word conversion), voice (preset and cloned voices), timing (word times for read-along) and the study model that writes and grades questions.   Files are kept per learner, and learner data in Postgres has an owner on every row.

The learning loop runs on top.   A check of about 12 mixed questions closes each chapter, and written answers are graded against the source passage.   What you miss feeds six ways to practice, and a spaced scheduler (FSRS) brings it back in one daily review across everything you read.   The daily review ran in the closed beta.   The fuller learning engine (profile, scheduler and mastery tracking) is still being built out.

Most of the product is ordinary software.   The AI does four jobs, and in each one I’ve decided in advance what it is allowed to do.

The main risk in a product like this is a confident wrong answer, such as a question with the wrong key or feedback that calls a correct explanation incorrect.   A learner might not notice and would end up studying the mistake.   So evaluation is a launch requirement.   Before any question type or feedback type ships, it has to pass a review set built from real chapters across subjects.   The set checks whether the answer is in the text, whether the key is right and whether the explanation helps.   The set now covers 400+ real chapters across 12 subjects, and the product has cleared all of its launch gates (see Results).   After launch I’ll track how often learners flag a question or dispute feedback, and fix patterns by question type.

Two more choices.   The library and the voice belong to the reader and are used only to serve that reader.   And I won’t make claims about learning outcomes until I’ve measured them.

Veritas isn’t publicly released, so these are early results from a waitlist, a closed beta and an offline evaluation, not results from a public launch.   I built the platform solo over 20 months: OCR import, voice cloning, read-along timing and an LLM study model running on a durable Postgres queue.

I’m running the launch in public stages.   A pre-release website explains the product, with a 2 minute 32 second film that walks through the whole loop in seven chapters, from “why we forget” to “one platform and why now,” and a waitlist for early access.   The first audience is students with an exam ahead, because the need is urgent and the habit is easiest to form when there’s a date.   Pricing will be announced before launch, and people on the waitlist hear first.   Public launch is planned for Q1 2027, after the first pilot.