Build a Second Brain Your Agent Remembers With Obsidian Wiki

You solve a hard problem, and three months later you solve it again from scratch in a different repo — the answer died in a chat log you will never find. obsidian-wiki, a 4k-star open-source project, points your agent at a folder and compiles what you learn into interconnected markdown you own.

Why This Skill Matters

The pattern comes from Andrej Karpathy's LLM Wiki gist, as the README credits: compile knowledge once and keep it current, instead of asking an LLM the same questions forever or re-running RAG every time. Every skill is a markdown file that any agent — Claude Code, Cursor, Codex, Windsurf, Gemini CLI, and more — reads and runs. No runtime, no API keys, no vendor. It compiles rather than accumulates: new knowledge merges into existing pages, contradictions get flagged, and a manifest tracks every ingested source so the second run processes only the delta.

Installation

pip install obsidian-wiki
obsidian-wiki setup --vault ~/brain

That also registers two session hooks: every new session starts with who you are and what's open, and ends by capturing anything reusable. Pass --no-hooks to skip them. uv tool install and pipx install work the same way. Prefer not to touch a terminal? Hand your agent this line from the README: https://github.com/Ar9av/obsidian-wiki — set up my wiki. On Claude Code you can install it as a plugin instead — no Python needed:

/plugin marketplace add Ar9av/obsidian-wiki
/plugin install obsidian-wiki@obsidian-wiki

The plugin ships the skills only — not the session hooks or the CLI — and its commands are namespaced, like /obsidian-wiki:wiki-ingest.

Real Workflow: Grow and Query the Vault

1. Feed it. /wiki-ingest ~/research ingests anything text-shaped — docs, PDFs, chat exports, transcripts, screenshots, URLs. /wiki-update distills the repo you are standing in, code-graph aware; /wiki-capture saves the current conversation; /wiki-history-ingest claude mines everything you have ever asked Claude.
2. Ask it. /wiki-query what do I know about rate limiting? answers with [[wikilink]] citations instead of vibes; /wiki-digest week recaps what you learned this week.
3. Keep it honest. /wiki-lint flags broken links, orphans, and contradictions; /wiki-dedup merges pages that mean the same thing; /wiki-status shows what is ingested and where the hubs are.

The README counts 39 skills in total, and every claim in the vault is tagged extracted, ^[inferred], or ^[ambiguous], so you can tell knowledge from guessing.

Tips

• Everything is plain markdown on disk: greppable, diffable, pushable to a private repo. Nothing leaves your machine.
• Open the vault in Obsidian's graph view and say color my graph — it tints nodes by tag, category, or visibility.
• Not ready to commit? The repo's examples/demo-vault/ is an 18-page sample vault on building reliable LLM agents.
• The whole graph exports to graph.json, GraphML, Neo4j Cypher, Postgres SQL, or a self-contained interactive graph.html.
• The README's own benchmark (Claude Sonnet, headless, a 38-page vault, two runs per question) reports structural questions answered 4.4x faster with the skills installed — 81s down to 19s — and correctness up from 44% to 83% at flat API cost. It also says to treat the exact percentages as indicative at that sample size, and the plain agent's failure mode is instructive on its own: it routed through index.md, which links to every page, and found a short path that meant nothing.

When Not to Use This

You want a hosted, multi-device notes service — this is a local folder by design; the README documents a browser extension for capturing pages and a Docker deployment for serving a vault over HTTP/MCP, but no sync service or mobile client. And if your notes never need to link, contradict, or accumulate, a plain folder of markdown files already does that job without 39 skills attached.

See the leaderboard for more skills.