Don't build a spaced repetition startup

I’m pausing full-time work on Rember, an AI-assisted spaced repetition system. We built a cleaner, opinionated product where you point to what you want to remember and get flashcards. I use it daily and love it, but we couldn’t turn it into a successful startup. I’ve lost conviction that a general‑purpose spaced repetition system can be venture‑scale: durable learning needs effort, habits are hard to form, and deciding what to remember is constant overhead. If you’re building in this space, I hope this blog post helps you pressure-test your approach. The walkthrough video below shows the latest version of Rember.

Rember

Anki has been the entry point to spaced repetition for millions of knowledge workers, from doctors to engineers. It’s free, fast, local-first, keyboard-centric, open and extensible, enabling the largest community of spaced repetition practitioners to tinker with and extend the system; it’s open-source and has been trusted for almost twenty years. But the UI is outdated, onboarding is steep, writing flashcards is clunky, sync feels fragile, the add-on ecosystem is brittle, and the product lacks opinionated guidance.

Andrea Vaccari and I set out to build a better spaced repetition system: cleaner, easier to use and opinionated. But ultimately we failed to make something people want and Rember didn’t get much traction.

Rember feels essential to me, I use it every day and no other spaced repetition system currently satisfies my personal needs. Therefore I’d love to keep maintaining and working on Rember as a side project in the foreseeable future. I don’t want to make promises, though. I simply lost the conviction that we can build a successful startup (see below), so I’m now exploring other directions for my career.

While learning something, I want to find good mnemonic prompts & open ended questions.

If I can’t, I want to write them myself and make them available for others (paid and free).

— @giacomo_ran, 2020-09-11

It was my pinned tweet for years after I posted it. This was my personal goal for Rember and what we were initially working towards. The first part of the tweet identifies a problem, the second part identifies a possible solution. The silver lining is that LLMs ended up solving the exact problem I was chasing: I can now generate flashcards for everything I read.

Even though Rember didn’t succeed as a company, I’ve grown enormously as an engineer, founder, and person. I’m grateful for the connections I made along the way. To all the people in the community who over the years warmly shared their enthusiasm for better spaced repetition tools, thank you.

Here’s a walkthrough of the latest version of Rember:

Challenges in building a spaced repetition system

In this part of the blog post I’ll share some of the challenges we faced while building Rember. Many of them became apparent only after many discussions with dozens of users, from spaced repetition experts to complete beginners. I hope they’re useful if you decide to build your own system.

For a deeper discussion of edtech product challenges, see:

Why Ed-tech Startups Don’t Scale by Gian Segato
Why is the EdTech Industry So Damn Soft? by Justin Skycak

Rember ran into three persistent sources of friction:

Desirable difficulties. Durable learning requires effort. While consumer products are optimized to minimize friction, learning requires preserving some of it. Remove the effort and you get edutainment; preserve it and you narrow your audience to serious learners.
Habit formation. Spaced repetition only works if you show up regularly. That means the product competes every day with everything else on your phone: Instagram, X, TikTok. When a system delivers value only through consistent use, it inherits the hard problem of creating that consistency.
Self-directed learning. Every time you learn something new, you have to ask yourself, “Should I turn this into flashcards?” That ongoing meta-decision adds friction and overhead to the learning process.

Many successful edtech products sidestep one or more of them:

ChatGPT as a tutor. Practice remains effortful and self-directed, you must bring material or specify a topic. But no daily habit is required to benefit, you can invoke it on demand.
Duolingo. You choose a target language at sign-up, then follow a predefined path, which greatly reduces self-direction. It relies heavily on gamification and aggressive notifications to support habit formation, leaning toward edutainment.
Math Academy. Like Duolingo, it provides a fixed curriculum, but it targets serious learners and intentionally maintains high difficulty.
Anki. In the general case, it faces all three sources of friction. For major use cases (language learning and medicine), the community heavily relies on pre-made shared decks, which reduces self-direction.

If you enter edtech without a clear plan to absorb, redirect, or remove these frictions, you risk shipping a clever tool that stalls at adoption. Rember set out to build a general learning tool, which meant confronting all three head‑on without a practical strategy to counter them. By not compromising on generality, we ended up with a tool destined to serve a passionate niche, but unlikely to escape it.

Challenges in the flashcard creation workflow

One of the biggest reasons people give up on spaced repetition is that creating flashcards by hand is tedious. LLMs partially solve this problem: they remove the need to manually draft every card yourself. But while they lower the barrier, they don’t make card creation effortless. You still need to decide what’s worth remembering, that decision is where much of the friction lives. In other words, AI helps, but it doesn’t erase the key bottleneck that keeps many people from building a sustainable spaced repetition practice.

Writing your own cards is valuable since it’s an effortful activity that forces you to deepen your understanding. But I find the common advice of “write your own cards” misplaced: you pay the opportunity cost of not reviewing all the cards you don’t make.

Products that generate flashcards with AI face a tight balancing act between:

Making the card creation process invisible. Ideally, you sign up, read something, and find fresh flashcards the next morning. If adding material requires conscious effort, most people stop unless highly motivated.
Generating cards which test exactly what the user cares about. Generate too many flashcards and you flood users with reviews. Generate too few flashcards and you might miss valuable information that they want to remember. You would need to somehow model prior knowledge and goals, because different readers often care about different aspects of the same text.

If you try to make the process fully invisible, you inevitably take away the user’s control; without that control, the flashcards often miss what actually matters to them.

You can think of flashcard generation as two separate tasks (terms borrowed from Ozzie Kirkby):

Targeting: identifying the ideas worth remembering
Construction: turning those ideas into flashcards

Construction is far easier to automate than targeting. Targeting is hard. A common solution is highlighting, but even when two people highlight the same sentence, they often care about different aspects (try highlighting just “Rome is the capital of Italy” on Rome’s Wikipedia page). Letting users attach brief comments to highlights can clarify intent.

With Rember, we flipped this. We let users specify, in natural language, exactly what they want to remember, and planned to add optional highlighting (we didn’t get there). This tradeoff works well for me personally, but it doesn’t feel like a 10x improvement over the traditional manual workflow; the mental overhead of deciding which ideas to turn into flashcards is still there.

Moreover, despite several iterations on our LLM prompt, frontier models still make subtle mistakes that add friction and erode trust. One idea is to treat the review session as an inbox: keep the useful cards, discard the rest. But that’s an inherently friction-filled workflow. Ultimately, I don’t think we’ve found the right balance with Rember; we need to go through several more UX iterations.

Generating cards with AI introduces additional minor issues, that didn’t show up as much in the traditional manual card creation workflow. If the user reads about the same ideas across different articles, the system likely ends up generating duplicate cards. Moreover, since adding many cards becomes easy, it would be beneficial to present concepts sequentially, by introducing prerequisites first. I discussed a potential technical solution to these and other problems in Content-aware Spaced Repetition.

AI makes flashcards cheaper to create, but it doesn’t remove an important part of the workflow: deciding what’s actually worth remembering. That targeting step is still a human judgment call, and it remains a source of friction. And even once you outsource construction to a model, you inherit new tradeoffs: the tension between giving users more control and keeping the workflow effortless, unpredictable outputs, and duplicate or poorly sequenced cards.

Further challenges

Even with a magical system where you study without thinking about spaced repetition and wake up to a perfect set of flashcards, several challenges remain unsolved.

The first is habit formation. Building a review routine still requires effort and discipline. After several conversations with smart knowledge-worker friends, my impression is that many do not see memory as their bottleneck, which limits their ability to learn and achieve their goals. By contrast, I and many others in the spaced repetition community see memory as a core problem. If you have “forgetting anxiety,” you are probably already using Anki; if you don’t, you are unlikely to build a review habit, no matter how seamless the card-creation workflow becomes. The difficulty is compounded by the fact that spaced repetition requires upfront effort for a delayed and uncertain reward. You often need to wait days or weeks before appreciating the improved retention, while competing with the instant gratification of social media or edutainment.

Another fundamental challenge is positioning. Niche apps can market themselves more effectively by focusing on a very specific use case. For example, if you need to pass the driver’s license theory exam, you can download a $5 app built solely to help you pass, with ready-made content, exercises, and progression. Alternatively, you can use a general-purpose spaced repetition app, where you must create your own cards, ensure they are correct, and cover all required topics. More broadly, people don’t wake up thinking “I wish I had a better system to remember things”. They worry about passing an exam or getting a promotion. Our product does not directly address these concerns. First you have to explain that spaced repetition is the best way to solve their problems, then you have to argue that your spaced repetition system is the best option among those available. It’s a tough sell.

Consumer apps like Duolingo show how powerful habit loops and progression systems can be, but they also highlight a structural gap. Once you pick a language, Duolingo never runs out of content. A general-purpose spaced repetition system is instead bounded by what the user is actively learning. If there is no fresh material, there is nothing new to review. To match Duolingo’s endless stream of content, you would have to expand the product into a tutor that curates and teaches new material, rather than only helping you remember what you have already encountered.

Beyond these strategic hurdles, there are practical frictions. Switching costs are unusually high. If I move from Notion to Obsidian and need an old note, I can briefly open Notion. Spaced repetition is different because you build a collection of flashcards to review over months or years. Unless a user is willing to review cards in multiple systems every day, switching means migrating the entire collection. That is why a robust import and export solution is a higher priority than in most other products. People develop preferences for how flashcards are written. Ideally, the system would adapt to each user’s style, but that is difficult to implement. As a corollary, the people who most want a flashcard generator, namely those already familiar with spaced repetition, are also the most likely to have strong preferences that the AI may not satisfy.

Conclusion

When I discovered Anki at university, I spent an extra hour after each one-hour class writing flashcards by hand. It clearly improved my quality of life: I could choose to remember anything I cared about. When we started working on Rember, we assumed that by removing this long, tedious part of the workflow, by solving flashcard creation, the potential market for spaced repetition would expand 100x. Instead, it looks more like a single-digit multiple. Not enough to justify a venture-backed company. Multiple sources of friction remain unaddressed, and they compound with each other. It is absolutely possible to build a solid lifestyle business around a spaced repetition product, but I have lost conviction that it can scale into a startup.

I don’t have prescriptions here. My hope is that these challenges help you sharpen a different approach, or decide to attack a different problem entirely.

I’ll be in San Francisco September 22-26. DM me on x dot com if you want to meet!