Obsidian-vault-agent recall

<Purpose> Run an interactive spaced repetition session: build a mixed queue of retrieval practice cards, inbox triage items, and a surprise cross-domain connection. For each card, show only the title first — forcing recall before revealing content (retrieval practice). Interleave domains to exploit the spacing and interleaving effects. Surface a non-obvious connection the user hasn't made yet. </Purpose>

<Use_When>

• User wants a review session ("what should I review?", "quiz me", "recall practice")
• User wants to resurface notes they haven't seen in a while
• User wants to triage their inbox
• User wants to discover connections across domains
• User asks "what do I remember about X?"
• Periodic knowledge maintenance (daily, weekly reviews) </Use_When>

<Do_Not_Use_When>

• User wants to deeply process a specific note (use /process)
• User wants vault structure analysis (use /vault-graph)
• User wants to synthesize across a cluster (use /synthesize)
• User is actively reading or adding a new source note </Do_Not_Use_When>

<Execution_Policy>

• INTERACTIVE — never skip ahead. Each card requires the user to respond before the answer is shown. The friction IS the learning.
• Never show note content before the user attempts recall
• Prefer notes with updated_date older than 14 days (spacing effect)
• MUST interleave domains — draw from at least 3 different topic folders
• Use MCP tools (search_notes, get_frontmatter) when available; fall back to Grep/Glob
• Session size: default 5 retrieval cards, 2 inbox items, 1 surprise connection (adjust to user's argument if provided)
• After each reveal, offer to update the note if it's stale </Execution_Policy>

<Steps>

Stage 1: BUILD THE QUEUE

1a. Retrieval Practice Cards (default 3–5)

Candidate selection using MCP:

search_notes(query="processing_status: processed OR processing_status: evergreen", limit=50)
search_notes(query="type: term OR type: note", limit=50)

Or fall back to Grep:

Grep(pattern="processing_status: (processed|evergreen)", path="notes/", glob="*.md")
Grep(pattern="^type: (term|note)$", path="notes/", glob="*.md")

Filter and rank candidates:

1. Prefer notes with

   updated_date

   older than 14 days — compute from today (2026-03-20)
2. Prefer notes with the

   [ ](#anki-card)

   anchor (marked for spaced repetition)
3. Prefer notes with

   processing_status: processed

   over

   evergreen

   (evergreen notes are maintained; processed ones need re-testing)
4. Enforce domain interleaving: collect the topic folder for each candidate (

   notes/ml/

   ,

   notes/psychology/

   ,

   notes/startup/

   , etc.) and ensure the final selection spans at least 3 different folders
5. If fewer than 3 domains are available, accept 2 minimum

Read the frontmatter to confirm

updated_date

and

type

:

get_frontmatter(path="notes/PATH/filename.md")

1b. Inbox Triage Items (1–2)

Find the oldest unprocessed source notes:

search_notes(query="processing_status: inbox", limit=20)

Or:

Grep(pattern="processing_status: inbox", path="notes/", glob="*.md")

Filter to types:

paper

,

post

,

book

,

lecture

,

course

. Sort by

created_date

ascending — oldest first. Pick 1–2.

1c. Surprise Connection (1)

Goal: find two notes from DIFFERENT domain folders that share a concept keyword but have NO wikilink between them.

Algorithm:

1. Pick a retrieval card already in the queue (it has topic keywords in its body)
2. Extract 2–3 key concept words from that note's title or content
3. Search a DIFFERENT domain folder for notes containing those keywords:

   Grep(pattern="KEYWORD", path="notes/OTHER_DOMAIN/", glob="*.md")
   

4. Check that the two notes do NOT link to each other:

   Grep(pattern="\\[\\[.*TITLE.*\\]\\]", path="notes/SOURCE_NOTE.md")
   

5. If no match, try a second keyword or a second candidate note
6. Present both notes with the shared concept as the "bridge"

Non-obvious pairings to prioritize (cross-domain > within-domain):

• ml/ ↔ psychology/ (learning algorithms ↔ cognitive science)
• startup/ ↔ design/ (product strategy ↔ UX)
• finance/ ↔ ml/ (optimization ↔ market dynamics)
• psychology/ ↔ startup/ (behavioral economics ↔ business)
• computer science/ ↔ logic/ (formal systems ↔ reasoning)

Stage 2: PRESENT QUEUE OVERVIEW

Before starting, show the session plan:

Review session — [N] cards queued

Retrieval practice ([count]):
  1. [Note title] — [domain]
  2. [Note title] — [domain]
  ...

Inbox triage ([count]):
  • [Note title] ([type], [age] days old)
  ...

Surprise connection: [teaser — e.g., "a bridge between ML and psychology"]

Ready? I'll show you one card at a time. Type anything to start, or "skip [N]"
to skip a card.

Stage 3: RUN THE SESSION

For each retrieval practice card:

Turn 1 — Prompt recall:

─────────────────────────────────────
Card [N/total] · [domain folder]

[[ (Type) Note Title ]]

What do you remember about this concept?
(Type your answer, or "skip" to reveal)
─────────────────────────────────────

Wait for user response.

Turn 2 — Reveal: Read the full note content:

read_note(path="notes/PATH/filename.md")

Or:

Read(file_path="/absolute/path/to/note.md")

Show the note content (truncate at 400 words if very long — show key sections).

Then ask:

How did you do? Is this note still accurate?
[u] Update note  [n] Next card  [e] Evergreen (promote status)

If user wants to update: help them edit inline, then update

updated_date

in frontmatter using

update_frontmatter

.

If user types "e" or "evergreen": update

processing_status

to

evergreen

.

For each inbox triage item:

─────────────────────────────────────
Inbox triage · [type] · [age] days old

[[ (Type) Note Title ]]

Read the note, show the first 5–7 bullets (or Abstract section for papers).

Then ask:

Worth processing deeply, or archive?
[p] Process (I'll run /process next)  [a] Archive  [s] Skip for now

• p

  → note it as a recommendation; optionally update

  processing_status: processing

• a

  → update

  processing_status: archived

• s

  → leave unchanged

For the surprise connection:

─────────────────────────────────────
Surprise connection

These two notes share the concept "[KEYWORD]" but don't link to each other:

[[ (Type) Note A ]] (domain: X)
[[ (Type) Note B ]] (domain: Y)

[Brief excerpt from each showing the shared concept]

Do you see a meaningful connection? What would the link say?
─────────────────────────────────────

Wait for user response. If they see a connection:

• Help them decide which note gets the wikilink (or both)
• Offer to add the link:

  [[  (Type) Note B  ]]

  into Note A's Related links section
• Update

  updated_date

  on both notes

Stage 4: SESSION SUMMARY

─────────────────────────────────────
Session complete

Reviewed: [N] cards
Updated: [N] notes
Promoted to evergreen: [list]
Archived: [list]
Queued for /process: [list]
New connections made: [count]

Oldest unreviewed term: [title] ([N] days)
─────────────────────────────────────

Optionally suggest: "Run /process on [oldest inbox item] next."

</Steps> <Examples> <Good> User: "review session" 1. Build queue: 4 term notes (ml/, psychology/, startup/, finance/), 2 inbox papers, 1 surprise 2. Show overview → user starts 3. Card 1: "[[ (Term) Attention Mechanism ]]" — user recalls, sees note, updates a stale bullet 4. Card 2: "[[ (Term) Loss Aversion ]]" — user blanks, reads note, promotes to evergreen 5. Inbox: "(Paper) Chinchilla Scaling Laws" — 60 days old → user says archive 6. Surprise: "(Term) Gradient Descent" ↔ "(Term) Operant Conditioning" — shared keyword "reward signal" → user says "oh, both use incremental feedback to converge on a target behavior" → adds link 7. Summary: 6 reviewed, 2 updated, 1 evergreen, 1 archived, 1 connection made </Good> <Good> User: "quiz me on 3 cards" → Session with 3 retrieval cards (not 5), 1 inbox item, 1 surprise connection </Good> <Bad> Showing note content immediately without prompting recall first. → This destroys the retrieval practice effect. ALWAYS show title first, wait for response. </Bad> <Bad> Picking 4 cards all from notes/ml/. → Violates interleaving. Domains must be mixed. If ml/ only has enough candidates, pull from any other folder before repeating a domain. </Bad> </Examples>

<Escalation_And_Stop_Conditions>

• Not enough processed notes (<3): Inform user, offer to run /process first on top inbox items
• All notes recently reviewed (updated within 7 days): Report "vault is fresh — nothing due for spacing"
• No inbox items: Skip inbox triage stage, replace with an extra retrieval card
• Surprise connection not found after 3 attempts: Skip it, note "no unlinked cross-domain pairs found today"
• User types "stop" or "quit": End session immediately, show partial summary
• MCP unavailable: Fall back to Grep/Glob for all searches; behavior is identical </Escalation_And_Stop_Conditions>

$ARGUMENTS