Vault Health

mnemo:health — Vault Health Check & Analytics

Run a comprehensive health check on the Obsidian vault: orphans, broken links, missing sections, stale notes, and growth statistics.

Prerequisites & config

Obsidian must be open; obsidian CLI on PATH. Config at ~/.mnemo/config.json (required fields: vault, taxonomy, links_section) — schema in ${CLAUDE_PLUGIN_ROOT}/references/config-schema.md.

Workflow

Steps 1-4 run in parallel — single assistant message batching the independent CLI reads (orphans, unresolved, tag counts; Step 4 reuses Step 3's tag output, no extra call). These are independent queries against the same indexed vault, ~180ms total vs ~720ms sequential.

Step 0: claude-mem Sanity Check (optional, ~20ms)

Surface two common gotchas if claude-mem plugin is installed and enabled. If cascade.claude_mem.enabled is false, skip this section silently — many users intentionally disable claude-mem for CPU/RAM reasons.

bash "${CLAUDE_PLUGIN_ROOT}/scripts/check-cm-version.sh"
# Or from source: plugins/mnemo/scripts/check-cm-version.sh

Script emits three lines: version: X, stale: N, path: .... Interpret:

• stale > 0 → warn: "claude-mem has {stale} old version folder(s) cached. Restart all Claude windows — old Stop hooks point to a path that no longer exists."
• `version {const u=app.metadataCache.unresolvedLinks;const f={};Object.values(u).forEach(l=>Object.keys(l).forEach(t=>f[t]=(f[t]||0)+1));return JSON.stringify(Object.entries(f).sort((a,b)=>b[1]-a[1]).slice(0,10));})()" vault="{vault}"

A short name with many refs (e.g. `[[Diadoc]]` ×30) = create a hub note `Diadoc.md` → `[[MOC — …]]` so all those links resolve (alias doesn't work for bare links — by design).

### Step 3: Tag Distribution

```bash
obsidian tags counts sort=count vault="{vault}"

Show top 15 tags. Flag tags used only once (potential typos).

Step 4: Notes by Type

Reuse the obsidian tags counts output from Step 3 — do not call it again (the tag index is one query; Steps 3 and 4 are two views of the same result). From it, extract counts for taxonomy tags: #atom, #molecule, #source, #session, #moc. These correspond to config.taxonomy.*.tag values.

Total notes count:

obsidian files ext=md vault="{vault}" total

Step 5: Missing Links Section (batched grep — 3600x faster)

Do NOT loop obsidian read per file — on a 1000-note vault that's ~180s. Use a single filesystem grep against the vault directory.

VAULT_PATH=$(bash "${CLAUDE_PLUGIN_ROOT}/scripts/get-vault-path.sh" "{vault}")

# Single recursive grep -L: files NOT containing the links section heading.
# Filter to taxonomy-prefixed notes.
grep -rL --include="*.md" "{links_section}" "$VAULT_PATH" 2>/dev/null \
  | grep -E "(Atom|Molecule|Source|Session|MOC) — "

Measured on 999-note vault: ~49ms vs ~180s serial — 3600x speedup. Safe to run always.

Report notes missing the section.

Step 6: Bad Filenames (# in names → permanent orphans)

VAULT_PATH=$(bash "${CLAUDE_PLUGIN_ROOT}/scripts/get-vault-path.sh" "{vault}")
find "$VAULT_PATH" -name "*#*.md" -not -path "*/.obsidian/*" -not -path "*/.trash/*" 2>/dev/null | sed "s|$VAULT_PATH/||"

Files with # in the name are permanent orphans — [[Note #1]] parses as [[Note]] + heading anchor #1, so nothing resolves to them (even existing links). Flag for rename (# → — or drop the #). Same for . mid-name (breaks CLI create). See ${CLAUDE_PLUGIN_ROOT}/references/tool-routing.md (naming rules).

Step 7: Review Candidates (content-staleness, type-aware)

A temporal signal, distinct from orphans (Step 1, which is structural): notes untouched longer than the threshold for their type are candidates for a re-read. Threshold precedence: per-note ttl: → review.staleDays. → review.staleDays.default → 30 (legacy). Age is measured from the newest of date or reviewed — so stamping reviewed: {today} on a still-valid note resets its clock. That snooze is what stops a stale list from rotting into guilt-debt (the canonical failure mode of review dates — see Gotchas).

VAULT_PATH=$(bash "${CLAUDE_PLUGIN_ROOT}/scripts/get-vault-path.sh" "{vault}")
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/review-candidates.py" "$VAULT_PATH" --limit 30
# Or from source: plugins/mnemo/scripts/review-candidates.py

Output: CANDIDATES\t{n}, then THRESHOLDS\t{json}, then one tab-separated row per note ({overdue_days} {type} {anchor_date} {anchor_src} {threshold_days} {relpath}), most-overdue first. {threshold_days} is the budget actually applied to that note (per-note ttl: if set, else its type's staleDays, else default) — show it as (type, Nd budget), never invent a ttl. Pure filesystem — independent of the obsidian CLI graph cache (no lag/lie risk). A missing review config section reproduces the legacy uniform 30-day behavior, so this is safe before any config migration.

Don't AND this with backlinks — a well-linked note can still hold outdated claims. Report candidates on their own; cross-reference Step 1 yourself if you specifically want "old and orphaned."

Step 7.5: Content Lint (optional deep pass — gated by review.lint.enabled)

Steps 1-7 are cheap and structural. This is the content pass — Karpathy's "lint": actually re-read the notes to judge whether claims have rotted, instead of trusting the calendar. Together the checks cover his four: orphans (Step 1) + concepts-mentioned-but-no-page (Step 2 unresolved links) + stale claims (Step 7) + contradictions (here). Gated by config.json → review.lint.enabled (default false) because it reads note bodies and costs tokens — skip this step silently when disabled.

When enabled, take the top review.lint.maxCandidates (default 15) candidates from Step 7 and have them read & judged on the model set by config.json → review.lint.model (default haiku; sonnet/opus for higher-quality verdicts). This health skill itself runs as a haiku fork and cannot upgrade its own model, so:

• If review.lint.model is haiku (or unset) → do the lint inline in this fork.
• Otherwise → spawn one subagent on model: {review.lint.model} (Claude Code: Task tool, subagent_type: Explore or general; Codex: spawn_agent) that reads the candidate note bodies in one batched pass (filesystem read, not one obsidian read per file) and returns the verdicts. Keeping the cheap Steps 1-7 on haiku while the lint runs on opus is the whole point of the split.
• Report the subagent's verdicts verbatim — never assume "all still-valid". The fork only aggregates what the subagent returned; it does not re-judge. The Step 9 Content-lint block and its count MUST reflect the subagent's actual breakdown: if the subagent returned 13 still-valid + 2 update-needed out of 15, the report says 13 still-valid, 2 update-needed, NOT 15 still-valid. Defaulting the count to the candidate total (as if everything passed) is a reporting bug — wait for the subagent's verdicts before writing the report.

Emit a verdict per candidate:

• still-valid → close the loop: stamp reviewed: {today} into the note's frontmatter via mcp__obsidian__str_replace — anchor on the frontmatter line inside the leading --- block (not a date:/reviewed: mention in the body): replace the existing reviewed: value, or if absent insert after date: (date: {d} → date: {d}\nreviewed: {today}). A confirmed-valid note then stops resurfacing without a manual edit. This auto-stamp is on by default (config.json → review.lint.autoStampReviewed, default true); only if the user set it to false do you just recommend the stamp and write nothing. When the lint runs in a spawned subagent (model ≠ haiku), that subagent does the stamping — it already holds the verdict and the note path. If a stamp write fails (Obsidian offline, or mcp__obsidian__str_replace unavailable in the subagent context), don't drop it silently — collect those note paths and surface them under the Content lint report block so the user can stamp them manually.
• update-needed → one line on what specifically looks outdated.
• contradicts [[Other Note]] → name the conflicting note; flag the older one against the newer.

Verdicts are triage, not truth — on haiku especially, expect false positives; even on opus, surface them as questions. The only write health ever performs is the reviewed: auto-stamp above, and only ever on a still-valid verdict — never content, never on update-needed/contradicts. It fires only inside this lint pass (which is itself off unless review.lint.enabled is true), and can be turned back to suggest-only with autoStampReviewed: false. The user stays in control.

Step 8: Top Hubs

Enumerate the MOC notes (by the moc taxonomy prefix), then count backlinks for each:

VAULT_PATH=$(bash "${CLAUDE_PLUGIN_ROOT}/scripts/get-vault-path.sh" "{vault}")
find "$VAULT_PATH" -name "{moc_prefix}*.md" -not -path "*/.obsidian/*" -not -path "*/.trash/*" 2>/dev/null | sed "s|.*/||;s|\.md$||"   # MOC names

For each enumerated MOC name, count backlinks:

obsidian backlinks file="{moc_name}" vault="{vault}"

Sort by count, show top 5. Keep the enumerated MOC-name list — Step 8.5 reuses it to find topics that have no MOC.

Step 8.5: Research-Gap Candidates (report-only — where the vault wants to grow)

Karpathy's lint also proposes what to research next. Steps 1-8 already collected the raw signal — turn it into growth suggestions without any new CLI calls (reuse Step 2's obsidian eval top-10 unresolved targets, Step 3/4's tag counts, and the MOC-name list enumerated in Step 8). Two cheap, computable gap types:

• Topic cluster with no MOC — a non-taxonomy topic tag with ≥5 notes (Milo's "mental squeeze point", the same trigger config-schema.md uses for creating a MOC) whose {moc_prefix}{Topic} is absent from Step 8's enumerated MOC-name list. Suggest: "12 notes tagged #auth, no MOC → create MOC — Auth?"
• Recurring external with no Source note — a top unresolved target from Step 2's obsidian eval list (the authoritative metadataCache top-10, not the CLI unresolved output) cited ≥5 times that reads like an external tool/paper/vendor (not a short project name, which instead wants a hub note) and has no Source — … note. Suggest: "[[LangGraph]] cited ×9, no Source note → capture one?"

These are suggestions, never auto-created (same non-destructive stance as the rest of health). Skip a type silently when it yields nothing. Do not web-search to fill the gap — mnemo maintains a human-authored vault: it points at the gap, the user decides whether to fill it. (This is the on-philosophy half of Karpathy's "suggest new article candidates"; the auto-web-imputation half is deliberately omitted.)

Step 9: Output Report

📊 Vault Health Report ({date})

⚠️ claude-mem: {warning or "v12.3.9, clean"}

Total: {N} notes
  Atoms: {N} | Molecules: {N} | Sources: {N}
  Sessions: {N} | MOCs: {N} | Other: {N}

🔴 Orphans: {N}
  - Note Name 1
  - Note Name 2

🟡 Missing {links_section}: {N}
  - Note Name 1

🚫 Bad filenames (`#`/`.`): {N} — permanent broken links, rename
  - Atom — Foo (PR #12)  → rename to "PR 12"

🔍 Top unresolved targets (missing hub notes?):
  1. [[Diadoc]] ×34 → create hub note?
  2. [[Python]] ×28

🔗 Unresolved wikilinks: {N} total
📏 Tags: {N} total, {N} used once

🏆 Top-5 Hubs (most backlinks):
  1. MOC — Security (34)
  2. MOC — AI ML Tools (28)
  ...

💤 Review candidates (stale by type-aware age): {N}
  - Atom — API X gotcha — 45d overdue (atom, 60d budget)
  - Source — vendor API pricing — 35d overdue (source, 180d budget)
  (snooze a still-valid note: add `reviewed: {today}` to its frontmatter)

🔬 Content lint: {N judged} — {S} still-valid, {U} update-needed, {C} contradicts   ← only when review.lint.enabled
  (counts MUST equal the lint's actual verdicts — never default to all-still-valid; see Step 7.5)
  - Atom — API X gotcha → UPDATE-NEEDED: superseded by [[Atom — API X v2]]
  - Source — vendor API pricing → still-valid (stamped reviewed: {today})

🌱 Research-gap candidates (where the vault wants to grow): {N}
  - #auth ×12 notes, no MOC → create MOC — Auth?
  - [[LangGraph]] ×9, no Source note → capture one?

🧠 Claude memory/ index: {KB}KB / {lines} lines {✅ lean | ⚠️ bloated → autodream}

Omit the 🔬 Content lint block entirely when review.lint.enabled is false. Omit the 🌱 Research-gap candidates block when Step 8.5 found nothing. The still-valid line above shows the default (autoStampReviewed: true — the note was stamped); with autoStampReviewed: false render it as → still-valid (recommend reviewed: {today}) instead, since nothing was written.

Skip the ⚠️ claude-mem line entirely if Step 0 found nothing to warn about.

Step 10: Claude memory/ index health (autodream check)

Separate from the Obsidian vault, Claude Code keeps an always-loaded index at memory/MEMORY.md. Claude Code auto-memory hard-truncates this index at ~24.4 KB on load — beyond that, trailing rows silently vanish from Claude's context. So warn early (before the cliff), not at some lax size. Threshold is configurable via config.json → memory.indexWarnKB (default 22):

WARN=$(python3 -c "import json,os;print(json.load(open(os.path.expanduser('~/.mnemo/config.json'))).get('memory',{}).get('indexWarnKB',22))" 2>/dev/null || echo 22)
for f in "$HOME"/.claude/projects/-*/memory/MEMORY.md; do
  [ -f "$f" ] || continue
  kb=$(( $(wc -c ${WARN}KB warn (auto-memory truncates ~24.4KB) → run autodream (move sessions → MEMORY-archive-index.md, target <20KB)"
done

If flagged → recommend autodream (memory consolidation): slim the index into topic files + MEMORY-archive-index.md, no loss. Procedure: ~/.claude/memory/autodream-principles.md. This is the only memory/ check here — vault-health otherwise audits Obsidian, not Claude's memory/.

Gotchas

Common failures in ${CLAUDE_PLUGIN_ROOT}/references/gotchas.md. Skill-specific rules:

• obsidian orphans may return empty on small vaults — this is OK, not an error.
• Reference notes (taxonomy docs, templates) aren't orphans even if few backlinks — they're meant to be lookups.
• Ghost notes (unresolved wikilinks) are a feature, not a bug — they enable entity discovery. Don't flag on raw count; instead surface the top targets (Step 2 eval) — frequent ones = missing hub notes (actionable).
• CLI graph queries cache & can lie — orphans/unresolved/backlinks lag writes and have shown a note as resolved AND broken at once. For critical checks use obsidian eval on metadataCache (see ${CLAUDE_PLUGIN_ROOT}/references/gotchas.md). Treat counts as advisory if notes were created this session.
• Do not auto-fix anything — only report. User decides what to clean up. The one exception is review.lint.autoStampReviewed (default true): it lets the content-lint stamp reviewed: {today} on a still-valid note (Step 7.5) to close the snooze loop. That is the sole frontmatter write health can make — only the reviewed: field of a confirmed-valid note, never content, never anything else. It fires only when the content lint itself is enabled (review.lint.enabled, default false), so a default install still writes nothing; set autoStampReviewed: false to keep the lint suggest-only.
• Step 5 uses filesystem grep (~3600x faster than per-file reads — 49ms vs 180s on a 999-note vault) — safe on any vault size.
• Review candidates (Step 7) are temporal, not structural — don't conflate with orphans. A note can be both, either, or neither. The script is age-only by design (cheap, no graph dependency).
• Content-lint verdicts (Step 7.5) are triage-grade — t

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jojoprison
• Source: jojoprison/mnemo
• License: MIT

Install and usage instructions live in the source repository linked above.