Personal-content re-surface decision math

What this skill does

Given (a) a corpus of saved personal content, (b) a static user profile, and (c) the user's recent activity, decide which N items to re-surface today and why. Output is a ranked list with one-line reasons. Does NOT define the UX (push notif vs daily digest vs dashboard widget) — that lives in product code.

Inputs

Output

[
  { "item_id": "yt_abc123", "score": 0.87, "reason": "topic match with current focus + 14d since last surface" },
  { "item_id": "wa_xyz", "score": 0.81, "reason": "high useful_score + 60d dormant" }
]

Rank descending. Cap at N (default 5). Each item carries a one-line reason for transparency.

Three algorithms — pick at build time

1. SM-2 (classic Anki) — day-1, no personalization needed

Decay-only. Each item has interval and ease_factor. When marked useful, interval grows. Re-surface when now > last_surfaced_at + interval. Zero LLM in the loop. Cheapest compute.

Use when: MVP day-1, corpus <100 items, no feedback signal yet.

2. FSRS (Free Spaced Repetition Scheduler) — modern, accurate

Three latent variables per item: difficulty, stability, retrievability. Updates via maximum-likelihood estimation. Outperforms SM-2 in benchmarks.

Use when: corpus >500 items AND ≥30 useful/not-useful events accumulated to train the model.

3. Custom context-aware hybrid — the Connecting-Dots moat

score = 0.5 × time_decay + 0.3 × relevance_to_recent_activity + 0.2 × static_profile_match

• time_decay: forgetting curve, sigmoid over days-since-surface
• relevance_to_recent_activity: cosine sim between item embedding and embedding of last 7d activity
• static_profile_match: cosine sim between item embedding and embedding of user profile

Use when: corpus >100 items AND user has authored a meaningful profile. The recommended path for Connecting Dots specifically — pure spaced-rep misses the "this is suddenly relevant" signal that recent activity unlocks.

See references/algorithms.md for exact formulas, half-life tuning, and edge cases.

Decision matrix

Common pitfalls

• Cold start: day-1 has no feedback. Default to SM-2 with ease=2.5, interval=1. Don't attempt FSRS.
• Recency bias: if only dynamic activity drives the score, old gold gets buried. The time_decay term fights this.
• Echo chamber: if relevance_to_recent_activity weights too high, the user only sees more-of-the-same. Cap that term at 0.5; force diversity sampling for top-N.
• Surface fatigue: enforce min_days_between_surface = 7 regardless of score — same item every day is annoying.
• Feedback bootstrap: cold-start the useful_score to 0.5 (neutral), update on user signal. Never let it drop below 0.1 — that's bandit-style starvation.

What this skill does NOT do

• Pull content from source platforms — that's baoyu-youtube-transcript, integrate-whatsapp, superpowers-chrome:browsing, etc.
• Store content — that's obsidian-vault, memory-router, or supabase.
• Extract entities/topics from content — that's ner-content-pipeline.
• Render the surface UX — that's product code (Obsidian dashboard, web app, push notif daemon).
• Build the RAG pipeline that answers queries — that's rag-patterns.

It only answers: "Given everything you have, what should I look at today?"