Obsidian Curator — Development skill for Claude Code

curator-agent

**The 4th layer of the Claude + Obsidian memory stack.**

A knowledge graph that grows without curation stops being a brain and becomes a landfill. `curator-agent` is the nightly maintenance pass that keeps retrieval accurate as your vault scales. Three jobs, all dry-run by default:

• rename — detect timestamp/category-shaped note titles, propose claim-shaped rewrites (the prose-as-title rule). Fixes wikilinks vault-wide on apply.
• link — detect notes making related claims that aren't yet linked, propose wikilinks between them.
• prune — surface notes untouched 30+ days with zero inlinks. Hub-aware: never flags your Maps of Content.

Stdlib-only Python. Zero dependencies. Runs in well under a second on a 200-note vault.

Why this exists

The original 3-layer memory stack ([article](https://x.com/nykdotdev/status/2030904887186514336)) was: session memory (CLAUDE.md + auto-memory), knowledge graph (Obsidian + MCP), ingestion pipeline (brain-ingest).

After 60 days in production, Layer 2 drifted. The ingestion pipeline fed the graph faster than it could be curated, and retrieval precision decayed. The fix wasn't a smarter model — it was a 4th layer that maintains the graph automatically.

Install

git clone https://github.com/0xNyk/obsidian-curator
cd obsidian-curator
# no pip install needed — stdlib only
python3 curator/curator_agent.py --self-test

Usage

# Dry-run report (default — never mutates anything)
python3 curator/curator_agent.py --vault ~/obsidian/your-vault

# Full report with link proposals + prune candidates, written to your vault
python3 curator/curator_agent.py --vault ~/obsidian/your-vault \
  --propose-links --report-prune \
  --report ~/obsidian/your-vault/00-home/curator-log.md

# Apply rename proposals (fixes wikilinks vault-wide)
python3 curator/curator_agent.py --vault ~/obsidian/your-vault --apply-renames

The nightly cron

# Read the report with your morning coffee
0 4