Spaced repetition can allow for infinite recall (2022)

Original Article ↗ Hacker News Discussion ↗

Tools and Algorithms in Practice

  • Many commenters focus on how to do spaced repetition: Anki, SuperMemo, FSRS, Mochi, MathAcademy, various language‑learning and kanji apps, and SRS‑integrated note tools.
  • FSRS (now in Anki) is praised for fitting a memory model to each user’s review history (card‑specific half‑life, target recall probability), but some users dislike its long intervals and find it miscalibrated, especially when prior learning happened outside Anki.
  • One proposed algorithm: model recall probability as (p = 2^{-\Delta/h}), fit (h) from recent reviews, and schedule reviews at configurable target probabilities (e.g., 90% vs 70%).
  • There is friction around UI/UX: some find Anki incomprehensible and seek simpler, Wanikani‑like tools; others say Anki is straightforward and life‑changing.
  • Open‑source FSRS libraries and bindings (TS, Rust) are appreciated, but documentation is considered weak. SuperMemo’s closed source and Pascal legacy attract both fascination and regret.

Debate over “Infinite Recall” and Memory Limits

  • Several people argue the article’s “infinite recall” claim is more math puzzle than psychology: it assumes unreal conditions (infinitely long life, oversimplified models), so its conclusions don’t map to reality.
  • Others defend such models as useful for revealing what isn’t the limiting factor and where a model breaks down.
  • Empirical counterpoint: decades of SuperMemo data suggest an upper bound of ~300k items; oral traditions show large but finite capacities. Overall capacity is seen as large but not literally unbounded.

When Spaced Repetition Helps (and When It Doesn’t)

  • Broad agreement: SRS is powerful for raw recall of discrete items—medical facts, legal cases, vocabulary, formulas, proofs—especially for exams and professional knowledge that must be instantly available.
  • Strong disagreement on scope:
    • Some say it’s “facts only,” with little relevance to math/engineering or deep understanding.
    • Others use SRS heavily for higher math, engineering, and conceptual material, claiming big benefits for later courses.
  • Many stress SRS should follow understanding: first grasp a concept in context, then encode it as cards to prevent forgetting.
  • Critics emphasize that real expertise comes from reading, discussion, experimentation, and problem‑solving; SRS is at best an optimization layer on top of that.

Language Learning and Other Domains

  • Language is the dominant use case: vocabulary, kanji, idioms, pitch accent. Some rely primarily on Anki (often with prebuilt decks); others say SRS fails beyond a few hundred words.
  • A major split:
    • Pro‑SRS: essential to make vocabulary “stick,” especially when combined with immersion; cloze deletions and mined sentences are recommended.
    • Skeptical: repeated real‑world exposure (reading, listening, conversations) and rich associations outperform flashcards; pure vocab decks often don’t transfer to fluent use.
  • Several describe hybrid systems: mining phrases from reading, integrating readers with SRS (auto‑marking words seen in context), and using SRS to gradually remove scaffolding (e.g., furigana).
  • For long texts, suggestions include chaining segments via index cards, cloze‑based text plugins, and incremental reading workflows (notably in SuperMemo and some Anki add‑ons).
  • Other domains mentioned: math (proof and problem cards), chess openings, driving‑test rules, and stock‑ticker knowledge unintentionally learned via internet “ambient repetition”.

Motivation, Boredom, and Habit Formation

  • A recurring practical problem: review backlogs (hundreds of due cards after a break) and burnout, especially when adding too many new cards per day.
  • Some argue that “boring but effective” is fine if it’s only 10–30 minutes daily; others counter that intolerable boredom kills compliance, so methods must be made engaging or replaced.
  • Suggested mitigations: fewer new cards, richer cards (audio, images, context), accepting that backlogs can be worked through gradually, and using SRS only for “high‑value” facts.
  • There’s also the view that most people should treat SRS as one small, disciplined habit—akin to going to the gym—complementing, not replacing, reading, practice, and immersion.

SRS in the Age of Google and LLMs

  • Some wonder if superhuman machine memory reduces the value of personal memorization.
  • Replies emphasize that what you can recall shapes your real‑time reasoning, problem‑framing, and conversation; you can’t effectively “just look up” things that never surface in your mind as relevant.
  • Consensus in the thread leans toward: search and LLMs lower the bar, but SRS remains highly valuable for those aiming at mastery or fields where fast, internalized recall matters.