I’ve been building a decision framework combining a few sources — Piotr Wozniak’s 20 Rules of Formulating Knowledge, Gwern’s “Spaced Repetition for Efficient Learning” essay, and Alex Mullen’s documented memory palace + Anki approach — and I want the community to stress-test it. It’s meant to apply to any new piece of knowledge, not just deducible facts — that was just my worked example, and I realize now it made the framework look narrower than it is.
The general framework, in order:
- Retention horizon: Is this indefinite (career-relevant) knowledge, or a one-off/administrative fact (an exam room number, a temporary password)? Only indefinite knowledge proceeds — everything else stays out of Anki entirely.
- Gwern’s cost-benefit filter: Regardless of what kind of fact it is — arbitrary or derivable — will the lifetime cost of not knowing it (repeated lookup time, or the cost of re-deriving it every time) exceed the lifetime review cost of memorizing it? Gwern’s estimate: a well-retained card converges to roughly 1.8–5 minutes of total review time over a lifetime; his rule of thumb is “don’t bother with SRS if you need it sooner than 5 days or it’s worth less than 5 minutes.” This gate applies to every candidate fact, deducible or not.
- Is it derivable from more basic principles (Wozniak’s Rule 1 — don’t learn without understanding)? This step does not decide whether the fact enters Anki — that was already decided in step 2. It only decides what kind of backup the card carries for when you fail it:
- Derivable → the backup is a reasoning chain (a “Big Picture” field with the derivation)
- Not derivable → check step 4
- Is it an arbitrary set/list of 5+ items with no internal logic (Wozniak’s Rules 9-10), or at risk of interference with a similar item already in the deck (Rule 11)? If yes, the backup is a memory palace image (Mullen’s approach — locus + minimal image description, following his own note that heavy narrative descriptions become useless once the fact is intuitive). If no, it’s a plain fact with no backup needed beyond the answer itself.
Two worked examples to show it’s not deducible-only:
- sin(30°) = 1/2 — derivable (bisect an equilateral triangle → 30-60-90 triangle). Passes Gwern’s filter (used repeatedly, re-deriving mid-problem is costly). Backup = reasoning chain, no image. Also flagged for interference with cos(60°) = 1/2, handled by anchoring which axis each corresponds to within the same Big Picture field, not a separate image.
- A drug’s exact half-life (e.g., amiodarone ≈ 58 days) — not derivable from any pharmacokinetic principle, it’s a measured empirical value. Also passes Gwern’s filter for a practicing pharmacist. Since it risks interference with other drugs’ similar half-lives, it gets a memory palace image + locus as its backup, per Mullen’s economy-of-effort principle (he explicitly avoids over-encoding facts that are intuitive or logically recoverable, reserving palace images for arbitrary/dense material).
So the same three-source framework produces different backups (reasoning vs. image) depending on step 3-4, but the entry decision (step 1-2) is identical regardless of whether the fact turns out to be derivable or arbitrary.
Does this full sequencing hold up — particularly the claim that derivability only determines backup type, never entry eligibility? Or is there a case where a derivable fact should be excluded from Anki for a different reason than the cost-benefit filter already covers?