Your Obsidian vault, but it actually remembers things

What it does

obsidian-second-brain is a Python skill that plugs into four AI CLI tools—Claude Code, Codex, Gemini CLI, and OpenCode—and gives them 43 commands to read, write, and reason about your Obsidian vault. Instead of treating notes as a dump, it treats them as a living database: ingest a meeting transcript and it extracts decisions, people, and tasks into the right pages; feed it a YouTube URL and it updates existing notes rather than appending yet another file.

The interesting bit

The core twist is rewrite, don’t append. The README explicitly frames this as an evolution of Karpathy’s LLM Wiki pattern: new information updates existing pages, contradictions get auto-reconciled via /obsidian-reconcile, and /obsidian-synthesize writes new synthesis pages without being asked. The notes themselves are structured for LLM retrieval—## For future Claude preambles and frontmatter—making the vault machine-readable first, human-readable second. Four scheduled agents (nightly close, weekly review, contradiction sweep, vault-health check) run while you sleep.

Key highlights

• 43 slash commands covering ingestion (/obsidian-ingest for audio, images, URLs), research (/research, /research-deep, /x-read, /x-pulse), reflection (/obsidian-challenge searches your own vault for past failures to push back on decisions), and visualization (/obsidian-visualize generates canvas maps)
• Dual-track research: open-web via Perplexity/Grok and source-grounded via Gemini File Search API (/notebooklm), with explicit encouragement to run both and chase contradictions
• v0.10 adds /obsidian-architect: scans a codebase and writes maintained architecture notes into your vault, refreshable without clobbering your edits
• Key-less web research and Google Calendar integration in the latest release
• Self-rewriting notes with anti-hallucination guards and a write-time AI-first validator

Caveats

• The README is heavy on promise and light on implementation detail: no visible architecture diagram, no explanation of how scheduled agents actually run (cron? daemon? CLI trigger?), and the “43 commands” are listed but not documented in the excerpt provided
• Requires multiple external API keys (Perplexity, Grok, Gemini, YouTube Data API v3) for full functionality; “key-less” appears to refer to a subset of research features, not the whole system
• The “AI-first” note format means your vault becomes increasingly opaque to human browsing without the CLI layer

Verdict

Worth a look if you’re already living in Claude Code or Codex and your Obsidian vault has become a graveyard of orphaned daily notes. Skip it if you want your notes to remain hand-curated or if the idea of an LLM rewriting your past decisions while you sleep sounds more alarming than helpful.