Spaced Repetition for AI PMs: How to Memorize Technical Concepts That Stick
TL;DR
AI PMs who use spaced repetition retain 80–90% of technical vocabulary at six months. AI PMs who don't retain about 25%. Set up Anki or RemNote in 30 minutes, seed it with the 100 cards in this article, and review for 10 minutes a day. That's the whole system.
Why Spaced Repetition Beats Re-Reading
Re-reading the transformer paper for the fifth time feels productive. It isn't. Cognitive science research (Karpicke and Roediger, 2008; Dunlosky et al., 2013) shows active recall plus spaced intervals beats every other study technique by 2–3x on long-term retention. The interval doubles each time you get a card right; the card resets to one day each time you fail it. The algorithm does the optimization — you just show up.
Re-reading
Comfortable, low effort, near-zero retention after 30 days. Feels like learning, isn't.
Highlighting
Same problem. The act of highlighting doesn't move information into long-term memory.
Spaced repetition
Effortful, 10 min/day, 80%+ retention at 6 months. Compounds: cards you've owned for a year cost 30 seconds a month.
Teaching it back (Feynman)
The strongest single technique. Best paired with spaced repetition: SRS keeps the building blocks loaded; Feynman exercises wire them together.
Tool Choice: Anki vs RemNote
Two tools matter in 2026. Pick one and stop researching. The tool you actually use is infinitely better than the perfect tool you don't.
Anki
Best for: PMs who want raw power, scriptability, and a free desktop app. The standard since 2006.
Setup: Install AnkiDesktop (free) + AnkiMobile (paid, $25 once). Create a deck called 'AI PM'. Default settings are fine. Limit new cards to 10/day to avoid review pile-ups.
Weakness: UI is dated. Adding cards is slower than RemNote. Markdown support is third-party.
RemNote
Best for: PMs who take notes and want SRS built directly on top of those notes. Read a paper, highlight a sentence, turn it into a card in two keystrokes.
Setup: RemNote.com, free tier covers individual use. Use the {{cloze}} syntax inside notes to auto-generate cards. The note-to-card pipeline is the real value.
Weakness: Less mature scheduling algorithm than Anki. Web-only on free tier.
100 Starter Cards: The AI PM Core Deck
Don't build cards from scratch the first week. Seed your deck with these 100 prompts spread across six topics. Add your own cards from there as you read papers and ship features.
Transformers (15 cards)
What does Q, K, V stand for? · Why is attention O(n^2)? · What is positional encoding for? · Decoder-only vs encoder-decoder use cases.
Training & Fine-Tuning (15 cards)
What is RLHF in one sentence? · DPO vs RLHF? · When does LoRA fail? · What is constitutional AI?
Inference Economics (15 cards)
Define TTFT, TPS, throughput. · Why is decode memory-bound? · What does KV cache do? · When does speculative decoding lose?
RAG & Retrieval (20 cards)
BM25 vs dense retrieval tradeoffs. · Why does chunk size matter? · What is hybrid search? · When does reranking help?
Evals (15 cards)
What makes a golden dataset golden? · LLM-as-judge biases (name 3). · Pass@1 vs pass@k? · Why are public benchmarks misleading?
Production & Safety (20 cards)
Direct vs indirect prompt injection. · What does MCP standardize? · Three guardrail layers. · Define hallucination operationally.
Card-writing rule: one fact per card. "What is RAG?" is one card. "What are the three failure modes of naive RAG?" is three cards (one per failure mode), not one giant card. Atomic cards are the difference between a deck that works and one you abandon.
Steal Our Pre-Built AI PM Deck
Masterclass students get the full 400-card AI PM Anki deck on day one — built by a Salesforce Sr. Director PM and updated quarterly as the field shifts.
The Weekly Cadence
The system fails when reviews pile up, not when cards are too hard. Protect the cadence first; everything else follows.
Daily (10 minutes)
Open the app first thing — before email, before Slack. Clear all due reviews. Cap new cards at 10/day. If reviews exceed 50, suspend new cards until you catch up.
Weekly (30 minutes, Sunday)
Add 20–40 new cards from the past week's reading, papers, and PRDs. Delete or rewrite any card you've failed three times — it's poorly worded, not you.
Monthly (1 hour)
Audit the deck. Suspend cards that no longer matter. Reorganize tags. Export retention stats. Aim for 85%+ true retention on mature cards.
Quarterly (2 hours)
Heavy refactor: merge duplicates, rewrite verbose cards, split overloaded ones. The deck should shrink as much as it grows.
Retention Numbers You Should Hit
Anki and RemNote both report retention. Track it. The numbers below are what consistent AI PM users hit after 90 days. If yours are far below, you have a card-writing problem, not a discipline problem.
Mature card retention: 85–92%
Cards you've seen 5+ times. Below 85% means cards are too dense — split them. Above 95% means intervals are too short — increase ease.
Young card retention: 75–85%
Cards seen 1–4 times. Lower is normal; this is when the algorithm is calibrating. Don't panic at 70% in week one.
Daily review time: 8–15 minutes
For a 400-card deck. If reviews exceed 25 minutes/day, you're adding cards too fast or your cards are too long.
True retention at 6 months: 80%+
Measured by unprompted recall in conversation, not in-app stats. Test yourself by explaining a random card concept out loud once a week.