Skip to content

review-buckets.md

Latest commit

 

History

History
52 lines (35 loc) · 3.15 KB

review-buckets.md

File metadata and controls

52 lines (35 loc) · 3.15 KB
Raw

Review buckets

Roots (see the repo README for full citations): Ebbinghaus's forgetting curve (knowledge decays without retrieval); Karpicke & Roediger on retrieval practice (the act of recalling, not rereading, is what produces durable memory); Wozniak's SM-2 algorithm (the canonical adaptive-interval scheduler, here simplified to three buckets because the LLM-driven substrate makes date-math fragile).

Spaced retrieval for retention, using three buckets instead of date-based intervals. Buckets are simpler than SM-2 and robust — no date math, no fragile intervals, and items move based on observed performance, not a schedule.

The three buckets

  • new — items introduced this session or very recently. Not yet tested for retention.
  • learning — items the learner has seen and partially holds. Due for retrieval practice. Most drill activity happens here.
  • mastered — items the learner has retrieved correctly on multiple past occasions with no hint needed.

Movement rules

  • new → learning: first successful retrieval after introduction.
  • learning → mastered: two consecutive clean retrievals on separate sessions (not two in the same session — same-session retrieval overestimates retention).
  • mastered → learning: any failed retrieval on a mastered item. Also add a note on what was missed.
  • learning → learning (stays): retrieval with hint, or partial retrieval.
  • any → dropped: if the learner explicitly says they don't want to revisit this, or it has been demoted three times (the item is poorly specified or the underlying concept is missing — address that instead of drilling).

How review.md looks

A single markdown file with three sections. Example:

# Review queue

## new
- Frame clauses (ROWS vs RANGE) — introduced 2026-04-24

## learning
- PARTITION BY vs GROUP BY — hinted 2026-04-22, clean 2026-04-24 (one more clean moves to mastered)
- ROW_NUMBER tie behavior — clean 2026-04-20

## mastered
- Basic window function syntax — last retrieved 2026-04-18

Items are one-liners. The prose after the em-dash is a note for you — last status, date, any relevant detail. Keep it readable; this is not a database.

Picking items for drill

At session start:

  1. Look at learning first. Pick 2–3 items, preferring ones you haven't drilled in the last session or two.
  2. Add 1 item from mastered as a retention check. Rotate which mastered items get checked — do not always pick the same one.
  3. If learning is empty and new has items, those items are not yet due for drill — they were just introduced. Let them settle for a session before drilling.

What not to do

  • Do not drill new items. They haven't been learned yet — the first retrieval should come after a short gap, not immediately.
  • Do not move things to mastered aggressively. Two clean retrievals on separate sessions is the bar. Same-session success means the item is still in working memory, not long-term.
  • Do not let learning grow past ~15 items. If it's overflowing, the learner is being introduced to new material faster than they can consolidate. Slow down new intake; drill the backlog.