Intro — May 2026, spaced repetition entered the "AI auto-generation" era

Just in 2024, using Anki meant building cards by hand. As of May 2026 that premise is shaking. **AnkiHub AI card generation**, **RemNote AI**, **Quizlet Magic Notes**, and **Brainscape AI** now produce card batches from a single PDF or lecture recording. Meanwhile the algorithm itself evolved. After 35 years dominated by SM-2 (1988), **FSRS (Free Spaced Repetition Scheduler)** landed in Anki 23.10 in late 2023 and became the de facto default.

This post is not the surface question "which app should I use." Instead, it traces the **algorithmic lineage of spaced repetition, the tool ecosystem, Korean and Japanese local flows, and the limits of AI auto-generation** all at once. We touch real usage patterns from USMLE prep to Japanese kanji learning to bar exam.

Why memory apps still matter in 2026 — why they survived the LLM era

A common 2024 question was: "If LLMs answer everything, why memorize?" By May 2026 the answer is clear. **Fast recall is a domain LLMs cannot replace**. When a med student must recall a diagnostic algorithm within 5 seconds, when a simultaneous interpreter has to pull a word instantly, when a lawyer recalls precedent in court — an LLM call is too slow.

Also, **recertification obligations** have grown. US medical boards, CPA, bar association credentials all require re-exams every N years. As a tool for maintaining what you once learned, spaced repetition remains number one. And with LLMs generating cards automatically, 2026 is the year **the barrier to entry collapsed**.

Algorithmic lineage of spaced repetition — Leitner through FSRS

Algorithms fall into six generations.

1. **Leitner System (1972)** — Sebastian Leitner's physical card-box system. Cards move between 5 boxes. The intuitive model behind every digital tool.

2. **SM-2 (1988)** — Introduced by Piotr Wozniak in SuperMemo 2. Per-card "ease factor" plus interval calculation. **Anki's classic algorithm** and a de facto standard for 35 years.

3. **SM-15 / SM-17 / SM-18** — The commercial SuperMemo line. SuperMemo 15 (2011), 17 (2016), 18 (2019) iteratively evolved. Instead of per-card scalars they track memory traces via the **DSR (Difficulty/Stability/Retrievability)** triplet.

4. **HLR (Half-Life Regression, 2016)** — Published by Settles & Meeder at Duolingo. ML predicts the "half-life" of each word. Foundation of Duolingo's stack.

5. **FSRS (2023)** — Open-source ML algorithm published by Jarrett Ye. Trains the DSR model on external datasets (20M+ reviews). **Bundled as default in Anki 23.10**.

6. **FSRS-5 / FSRS-6 (2025~)** — Extended to 21 parameters. Personal fine-tuning kicks in around 10k reviews of your own data.

The biggest shift is the SM-2 to FSRS jump. Where SM-2 "expressed card difficulty with a single ease factor," FSRS uses **3 latent variables (D/S/R)** and ML training to predict more accurately. In 2025 Anki user telemetry, 80% had migrated to FSRS.

The Anki ecosystem — Damien Elmes' 25-year project

**Anki** means "memorization (暗記)" in Japanese. Damien Elmes created it in 2006 and has maintained it ever since. As of May 2026 **Anki 24.06.3** is the current stable build.

The ecosystem looks like this.

- **AnkiDesktop**: Windows/Mac/Linux. GPL-3.0 open source. Full feature set.

- **AnkiDroid**: Android. Free. Maintained by a separate community.

- **AnkiMobile**: iOS. Paid (USD 24.99 perpetual). Its revenue funds the whole project.

- **AnkiWeb**: Free cloud sync. Card review available in the browser.

- **AnkiHub**: Collaborative editing SaaS (USD 6/month). The standard for medical-student communities.

- **AnkiHub AnKing decks**: The biggest med-student deck for USMLE Step 1/2. 350k+ cards.

Anki's strength is **extensibility**. 1500+ registered add-ons let you bend almost any workflow.

Anki card types — Basic, Cloze, Image Occlusion

Three core card types cover 90% of usage.

- **Basic**: Front/back. The most basic.

- **Cloze deletion**: Fill in the blank. Form: "The capital of Korea is [...]". Most common in medical learning.

- **Image Occlusion Enhanced (IOE)**: Cover a region of an image and recall it. Essential for anatomy, neuroanatomy, and ECG learning.

A Cloze card example syntax looks like this.

The capital of {{c1::Korea}} is {{c2::Seoul}}.

Written this way, Anki auto-generates 2 cards. One hides "Korea," the other hides "Seoul."

FSRS in depth — the structure of the DSR model

FSRS represents a card's memory state with 3 latent variables.

- **D (Difficulty)**: The intrinsic difficulty of the card. Scale 1 to 10.

- **S (Stability)**: How stable the memory is, in days. Larger S = longer next interval.

- **R (Retrievability)**: Probability of recalling right now. 0 to 1.

The next review time targeting a retention level (default 90%) is estimated as follows.

next_interval = S * (R_target^(1/decay) - 1)

When the user grades a card with one of "Again/Hard/Good/Easy," D/S/R update. Anki's built-in FSRS parameter optimizer treats your review log as training data to fine-tune 21 parameters. Generally, beyond 10k reviews personalization becomes meaningful.

Source is on GitHub at `open-spaced-repetition/fsrs4anki`, with Python/Rust/JS ports. RemNote, Trezo, Memo, and several other apps adopted FSRS too.

AnkiHub & AnKing — the standard of medical-student learning

**AnkiHub** is a SaaS for collaboratively editing Anki cards. Its core value is twofold.

- **Live updates to officially maintained decks**: When one student fixes a card, all subscribers receive the same edit. A git-like model.

- **Community validation**: When medical guidelines change, med-student moderators update the cards.

The most famous deck is the **AnKing Step 1/Step 2 Deck**. A megadeck of 350k+ cards for USMLE prep. An unofficial estimate has it that 80% of US medical students use this deck as of May 2026. A 4-tier subscription (USD 6 to 30/month) funds it.

AnkiHub added **AI card generation** in 2025. Learners upload lecture-slide PDFs, AI drafts cards, and community validation merges them into the megadeck.

Anki add-ons — AI auto-generation & TTS

Popular AI-related add-ons as of May 2026 are as follows.

- **Anki Card Generator GPT**: Calls the ChatGPT API to auto-generate cards from text. PDF upload supported.

- **AwesomeTTS**: Auto-converts words to TTS audio. Backends include Google, Azure, Amazon Polly, and ElevenLabs. Essential for language learning.

- **Image Occlusion Enhanced**: Image-masking card type. The original IOE.

- **HyperTTS**: A reinforced ElevenLabs-integrated variant.

- **Review Heatmap**: Visualizes review activity like a GitHub contribution graph.

RemNote — notes and flashcards merged

**RemNote** promises "notes and flashcards combined." You build outlines like in Notion and embed flashcards inline within them. By targeting med students first, as of May 2026 it has reached around USD 2M ARR with the number-two position in med-school learning after Anki.

Core traits:

- **Inline flashcards**: One-line syntax like `Term :: Definition` produces a card inside the note.

- **RemNote AI**: STT on lecture audio auto-generates cards and summaries. Launched 2025.

- **FSRS adoption**: Switched the default from SM-2 to FSRS in 2024.

- **Daily Doc, Backlinks**: Roam-Research-style bidirectional links.

- **Equations, code blocks, diagrams**: Suited to technical learning too.

Pricing: Free / Pro (USD 12/month) / AI (USD 19/month) in three tiers.

SuperMemo 19 — Wozniak's original tool

**SuperMemo** is the original spaced-repetition system Piotr Wozniak began in 1985. As of May 2026, **SuperMemo 19** (Windows desktop) plus **SuperMemo Web** is the active line. Algorithm: SM-18.

Strengths:

- **Incremental Reading**: Partially read PDFs or wiki articles and auto-create cards. A feature since the 1990s with no real equivalent in other tools.

- **Theoretical depth**: Wozniak releases an algorithm white paper annually. The SM-18 DSR-model docs are academically citable.

Weaknesses:

- **Weak mobile**: SuperMemo Mobile (iOS/Android) exists but is feature-poor versus desktop.

- **UI still feels 1990s**: Steep learning curve.

- **Commercial**: Lifetime license USD 60 to 90.

For serious learners (researchers, polymaths, megadeck owners) it is still number one, but for the general user Anki is the easier entry.

Quizlet + Magic Notes + Q-Chat — old champion reborn by AI

**Quizlet** is a US flashcard tool created in 2005 by high-schooler Andrew Sutherland. Once valued at USD 1.8B, it was losing ground to Anki. From 2024 it relaunched as an AI learning platform with **Magic Notes** and **Q-Chat** (AI tutor).

- **Magic Notes**: Upload lecture notes and AI auto-generates flashcards, summaries, quizzes, and mind maps.

- **Q-Chat**: A card-driven conversational AI tutor. Asks Socratic-style questions.

- **Learn mode**: Adaptive learning — similar to Anki's FSRS but on a proprietary algorithm.

- **Free + Plus (USD 7.99/month)**: AI features require Plus.

A detailed EdTech market analysis is in a separate post (iter83). From an SR perspective, **Quizlet's SR algorithm is not as polished as Anki/FSRS**. But the UX and card-sharing library (500M+ decks) are unrivaled. The primary tool for high-schoolers and undergrads.

Brainscape — Confidence-Based Repetition (CBR)

**Brainscape** is a US company founded in 2010. Its algorithm differs from Anki/FSRS — it uses **CBR (Confidence-Based Repetition)**. The user rates a "1 to 5 confidence" immediately after seeing a card, and intervals adjust from that. Simple but quick to onboard.

- **AI Tutor**: Added a GPT-4-based tutor from 2025. Explains cards and unpacks hard concepts.

- **Smart Cards**: AI auto-generates cards from your notes.

- **Free + Pro (USD 9.99/month)**.

- **Class plan**: School-level licensing. Strong in the GMAT, MCAT, and MS1 markets.

Mochi.cards — a modern Markdown-first tool

**Mochi** is a modern flashcard app from Canadian solo developer Anthony Bullard. Its key differentiator is **Markdown-first**.

Korea Capital

What is the capital of Korea?

Seoul

You manage cards as Markdown files. Compared to Anki's SQLite DB, the card library is much easier to manage with GitHub.

- **FSRS adoption**: Since 2024.

- **AI card generation**: ChatGPT API integration.

- **Lite + Pro (USD 5/month)**.

- **First-class support for equations, code, tables, and LaTeX**.

Adoption is growing among technical learners (engineers, med students). Not as powerful as Anki but worth it if you want a cleaner UX.

Memrise — pivoted to language-only

**Memrise** was founded in 2010 by Ed Cooke. Originally a general SR tool, it pivoted in 2022 to **language-only**. The algorithm is a proprietary SR plus AI conversation practice.

- **MemBot**: AI conversation partner. GPT-4 based.

- **Native Speaker Videos**: Native-speaker clips for pronunciation and culture.

- **Free + Pro (USD 8.99/month)**.

No longer recommended outside language learning. **TinyCards (Duolingo, 2017 to 2020)** had a similar fate — Anki effectively monopolizes the general SR market.

Pimsleur, Duolingo, WaniKani, Bunpro — language-specific tools

Let us look at how SR shows up in language learning.

- **Duolingo**: Uses HLR (Half-Life Regression) to time word exposure. Gamification plus AI (Duolingo Max, GPT-4).

- **Pimsleur**: An audio course rooted in graduated-interval-recall since the 1960s. The original SR applied to listening and speaking.

- **WaniKani**: Japanese kanji (漢字) only. SR levels 1 to 9. **Tofugu** operates it. USD 9/month or USD 299 lifetime.

- **Bunpro**: Japanese grammar SR. Tracks JLPT N5 to N1 progress.

- **iKnow!**: A Japanese-learning SR operated by DMM. Number one in the Japanese domestic market.

These are not generic SR tools but bundles of **language-specific content plus SR**. Compared to free combos like Anki + Core 6k, content curation is the value-add.

Korean-learning tools — Classcard vs. Quizlet Korea

The Korean market is led by **Classcard**. Acquired and operated by NHN Edu in 2014. Dominant market share among middle and high schoolers.

- **Classcard Matching/Memorize/Recall modes**: Four-stage gamified learning.

- **AI auto card generation**: Auto-extracts vocab from English sentences.

- **Teacher-student collaboration**: Integrated with hagwon (cram-school) and public-school infrastructure.

- **Free + Premium (KRW 9,900/month)**.

Alternatives:

- **Quizlet Korea**: Quizlet's Korean market. Mainly university students.

- **Wordboard**: A solo-dev app. Strength is simplicity.

- **MEMRISE Korean**: For Korean learners (foreigners). Koreans do not use it for English learning.

University students and above tend to switch to a free combo like **Anki + Core 2k/6k Korean**.

Japanese-learning flow — Tangocho Maker, iKnow, WaniKani

The Japanese market is fragmented.

- **Tangocho Maker (単語帳メーカー)**: A free iOS/Android app. The standard for Japanese exam-prep students. Very simple UI.

- **iKnow!**: Operated by DMM. Content curation for JLPT prep. USD 9.99/month.

- **WaniKani**: For foreign learners of Japanese kanji. SR plus mnemonics.

- **Anki + Japanese megadecks**: Core 2k/6k, Tango N5 to N1, JLPT Tango. All free.

- **AnkiWeb Japanese add-ons**: Japanese morphological analysis (MeCab), auto-furigana.

The student market goes to Tangocho Maker, working adults to iKnow, foreign learners to WaniKani plus Anki. Classic decks like **Core 6k** have an unofficial estimate of having been used by 1M+ people since 2007.

Medical / professional-license learning — AnKing, Pixorize, Sketchy

SR tools in the professional market follow a separate track.

- **AnKing**: USMLE medical megadeck (covered above).

- **Pixorize**: Visual mnemonics for biochem and pharmacology. Combined with SR.

- **Sketchy**: Visual-storytelling-based medical learning. SR plus video.

- **UWorld**: Med-school plus other certification Q-Bank. Weak SR integration.

- **Kaplan**: Classic in-person prep plus digital cards.

- **AdaptiBar**: SR for the US bar exam (MBE).

- **BarMax**: Bar prep. Content authored by Harvard Law School faculty.

These are not just tools — they are **content plus SR infrastructure**. Pricing runs USD 200 to 2000/year, an order of magnitude above general tools.

Notion · Obsidian · LogSeq integration plugins

The trend of integrating PKM (Personal Knowledge Management) and SR is strong.

- **Notion Flashcards**: Exposes Notion database rows as cards. Several unofficial plugins.

- **Obsidian Spaced Repetition**: The popular plugin from st3v3nmw. Adopted FSRS. Reviews cards directly inside notes.

- **Obsidian to Anki**: Exports cards from Obsidian notes to Anki.

- **LogSeq spaced repetition**: A built-in LogSeq feature. Supports cloze syntax.

- **Roam Research Anki**: A Roam-to-Anki sync plugin.

The core thesis of the PKM + SR convergence is that "**the moment you make a card is the moment you write a note**." You do not have to launch Anki separately, so friction drops.

AI card auto-generation — from one PDF to a hundred cards

Let us look at the practical reality of AI auto-generation as of May 2026.

1. **Upload PDF/slides**: Lecture slides or textbook-chapter PDFs.

2. **AI extracts core propositions**: An LLM decomposes content into proposition-level chunks.

3. **Auto Cloze deletion**: Core terms become blanks.

4. **Auto image-occlusion detection**: OCR plus object detection masks label boxes.

5. **TTS-added audio**: Auto-generate pronunciation for words.

Representative services:

- **AnkiHub AI Generator** (USD 6 to 30/month)

- **RemNote AI** (USD 19/month)

- **Quizlet Magic Notes** (USD 7.99/month)

- **Mochi AI** (USD 5/month)

- **Brainscape AI Tutor** (USD 9.99/month)

- **Local solution**: With Ollama plus the Anki Card Generator GPT add-on, you can generate cards from a local GPT-OSS model.

That said, **the quality of AI-generated cards is still only 50 to 70% usable**. Per interviews with med-student users, 50% are usable as-is, 30% need edits, and 20% must be thrown away. The standard flow is AI generation as a first draft followed by a human review.

Browser extensions — Polar Bookshelf, Readwise to cards

The flow from "stuff you read" to cards is also being automated.

- **Polar Bookshelf**: PDF/EPUB reader plus auto-export from highlights to Anki. Open source.

- **Readwise**: A highlight-aggregation SaaS. Pulls from Kindle, Pocket, Twitter, etc., then Readwise Mastery and export to Anki, Mochi, or RemNote. Covered in the RSS/reader post too (iter102).

- **Hypothes.is**: Web-page annotation tool. A plugin converts annotations into Anki cards.

Their primary purpose is not SR; it is **highlight preservation with an SR option**. The flow: auto-create 30 essential cards from a single book.

Academic research — the scientific foundation of SR

Let us see why SR works through academic research.

- **Pavlik & Anderson (2008)**: A memory model based on the ACT-R cognitive architecture. The scholarly base for SR algorithms.

- **Settles & Meeder (2016)**: The Duolingo HLR paper. Published at ACL 2016. The starting point of ML plus SR.

- **Lindsey, Shroyer, Pashler, Mozer (2014)**: MOI (Memory Optimization through Interleaved practice). Adopted by Carnegie Learning.

- **Bjork's "Desirable Difficulty"**: The learning theory of Robert Bjork (UCLA). Scientific basis for SR.

- **Karpicke & Roediger (2008)**: The testing-effect research. Evidence that recall beats plain re-reading.

These studies show that SR is not a mere technique but a **crystallization of decades of cognitive psychology**.

Combination patterns — how real learners stack tools

Recommended stacks by learning objective.

- **Med school main course + USMLE**: AnkiHub (AnKing) + Anki Mobile + Image Occlusion Enhanced + AwesomeTTS.

- **Bar exam**: AdaptiBar + Anki + BarMax content.

- **Japanese N3 to N1**: WaniKani + Bunpro + Anki (Tango N3 to N1).

- **Korean cram-school student**: Classcard + the school's own content.

- **Researcher/polymath**: SuperMemo 19 Incremental Reading + supplementary Anki.

- **Technical learner**: Mochi + Obsidian Spaced Repetition + AI auto-generation.

- **General undergrad**: Quizlet Magic Notes + departmental card libraries.

Looking purely at SR efficiency Anki + FSRS wins, but **friction to start** and **content curation** are the real variables in tool choice.

Adoption roadmap — starting from zero

Stages for a first-time SR learner.

1. **Week 2**: Make about 50 cards in Anki Mobile or Quizlet and review them every day. Habit-formation is the number-one priority.

2. **Month 1**: Separate card categories (language, job knowledge, general knowledge). Measure review time.

3. **Month 3**: Migrate to FSRS (automatic in Anki 23.10+) and run the parameter optimizer once.

4. **Month 6**: Add one AI auto-generation tool. Try producing a card batch from one PDF.

5. **Year 1**: Image Occlusion, advanced Cloze syntax, combined image-plus-TTS cards.

6. **Year 2**: Complete a single megadeck in your domain. 10k cards or more.

A **daily 15-to-30-minute review habit** matters far more than the tool. Anki's average user reviews about 25 minutes a day.

Closing — May 2026, "SR is a habit, not an algorithm"

This post covered algorithms and tools in depth, but the final conclusion is simple. **90% of SR is the habit of opening the app every day**. Even if FSRS beats SM-2 by 10%, if you do not open it daily that becomes 0%.

AI auto-generation lowered the entry barrier. The 2024 excuse "making cards is annoying" no longer holds in 2026. We have reached the point where one PDF produces a card batch automatically.

Decide on a tool in about a week, and then forget the choice. The real work is the daily 25-minute review.

References

- Anki official site: https://apps.ankiweb.net/

- Anki Manual: https://docs.ankiweb.net/

- AnkiHub: https://www.ankihub.net/

- AnKing decks: https://www.ankihub.net/decks/

- FSRS GitHub: https://github.com/open-spaced-repetition/fsrs4anki

- FSRS Algorithm explainer: https://github.com/open-spaced-repetition/fsrs4anki/wiki

- SuperMemo: https://www.supermemo.com/en

- SuperMemo Algorithm SM-18 white paper: https://supermemo.guru/wiki/Algorithm_SM-18

- RemNote: https://www.remnote.com/

- Mochi.cards: https://mochi.cards/

- Brainscape: https://www.brainscape.com/

- Quizlet: https://quizlet.com/

- Memrise: https://www.memrise.com/

- WaniKani: https://www.wanikani.com/

- Bunpro: https://bunpro.jp/

- iKnow!: https://iknow.jp/

- Classcard: https://www.classcard.net/

- Tangocho Maker (App Store): https://apps.apple.com/jp/app/id403838621

- Pimsleur: https://www.pimsleur.com/

- Duolingo HLR paper (Settles & Meeder 2016): https://aclanthology.org/P16-1174/

- Pavlik & Anderson 2008 ACT-R memory model: https://act-r.psy.cmu.edu/wordpress/wp-content/uploads/2012/12/735JEPApplied.pdf

- Roediger & Karpicke testing effect: https://psychnet.wustl.edu/memory/wp-content/uploads/2018/04/Roediger-Karpicke-2006_PsychScience.pdf

- Obsidian Spaced Repetition plugin: https://github.com/st3v3nmw/obsidian-spaced-repetition

- Polar Bookshelf: https://getpolarized.io/

- Readwise: https://readwise.io/

- Image Occlusion Enhanced add-on: https://ankiweb.net/shared/info/1374772155

- AwesomeTTS add-on: https://ankiweb.net/shared/info/1436550454

- Damien Elmes blog: https://blog.ankiweb.net/