Mohamed-El-Sharqawy Agent Tutor — Review
Understanding decays. This skill runs recall-first review sessions and keeps the retention schedule honest.
Vault root: OBSIDIAN_VAULT env var, else learning/ in the current workspace. All content under <vault>/Learning/.
The schedule
Every lesson note carries a review: block in its frontmatter with the note's memory state:
review:
interval: 6 # current interval in days
ease: 2.5 # multiplier applied on a solid recall
due: 2026-08-28 # next review date
Scheduling is FSRS-inspired and adaptive — no fixed ladder and no ceiling. After each review, compute the next state:
| Verdict | Interval | Ease |
|---|---|---|
| Solid | round(interval × ease) |
+0.05, up to max_ease |
| Shaky but recoverable | round(interval × 1.2) |
−0.2, down to min_ease |
| Gone | reset to the initial interval | −0.5, down to min_ease; flag for re-study |
Defaults — overridable per learner with an optional review_policy: block in the learner profile:
review_policy:
scheduler: fsrs-inspired
initial_interval: 1 # days after first learning / after a lapse
ease: 2.5 # starting multiplier
min_ease: 1.3
max_ease: 3.5
max_interval: null # null = unbounded; set e.g. 365 to cap growth
fuzz: true # ±5% jitter on computed due dates so notes don't pile up on one day
Notes are never "done": a long interval just means the topic comes up rarely. Legacy notes whose review: field is a plain list ([+1d, +3d, +7d]) still work — take the last entry as the current interval with default ease, and migrate them to the block format at this review.
Running a review session
- Read the review queue from the dashboard:
- Markdown mode (default): read
Learning/Dashboard.md→ Up for review section (notes whose next review date has passed). - Html mode (
output_format.dashboard: htmlin the learner profile): read theagent-tutor-stateJSON island at the top ofLearning/Dashboard.html's<body>— itsdue_notes[]entries (note, due date, interval) are the queue. Map each entry back to its actual note file underLearning/<Subject>/notes/(island labels are display names); the thinDashboard.mdhub is a signpost — never parse it. NoDashboard.htmlyet → the fallback below applies. - If the dashboard looks stale in either mode, scan
Learning/<Subject>/notes/*/frontmatter directly — notereview:frontmatter is always the scheduling authority.
- Markdown mode (default): read
- Recall first, always. For each note: ask the user to explain the topic from memory before showing anything. ("Explain closures to me as if I'd never heard of them.")
- Judge the recall against the note's key takeaways, then apply the schedule table:
- Solid → apply the solid row.
- Shaky but recoverable → show the key takeaways, have them re-explain; apply the shaky row only if the second attempt is clean, otherwise treat as gone.
- Gone → mark for re-study: re-open the lesson, re-teach the gaps, apply the gone row (reset to the initial interval).
- Interleave: mix topics from different phases/subjects in one session — interleaving is the point, don't review one phase in isolation.
- End with a mixed mini-quiz (5–8 questions spanning everything reviewed today).
- If an interactive quiz tool is available (e.g. pi's
quiz), use it. - Otherwise run a chat quiz: one question at a time, wait for the answer, explain why the right option is right, track the score, report pass/fail vs 70% at the end.
- If an interactive quiz tool is available (e.g. pi's
- Update each note's
review:frontmatter, the Dashboard review queue, and append a session log entry (typereview) toLearning/<Subject>/logs/YYYY-MM-DD.mdwith per-topic retention verdicts.
Honesty rules
- Judge recall against the note's key takeaways, not against "close enough".
- Tell the user their actual retention state — "you've forgotten X entirely" is useful information.
- If more than half of a subject's notes fail review, say it plainly and propose a re-teach plan rather than patching holes.
Dashboard maintenance
Keep the Up for review section current: one bullet per due note, sorted by due date:
- [[notes/03-closures]] — +7d interval, due 2026-08-24
Remove bullets after the review is logged. This list is what the tutor reads at session start, so it must be true.
Html mode (output_format.dashboard: html in the learner profile): the queue lives in Dashboard.html — the Up for review section and the due_notes[] of the island read in step 1. Update it at the end of the session by regenerating the html dashboard whole — a fresh island, overview and subject focus pages together, plus the thin Dashboard.md hub's counters, exactly the update-the-Dashboard moment every session ends with; never hand-patch the html.