Obsidian Wiki is a set of 39 agent skills that compile what you learn into an interlinked markdown wiki in an Obsidian vault. It follows Andrej Karpathy's LLM Wiki pattern: knowledge is compiled once and kept current, instead of asking an LLM the same questions again or re-running RAG each time. Each skill is a markdown file that coding agents such as Claude Code, Cursor, Codex, Windsurf, and Gemini CLI read and run, with no separate runtime or API key.
You feed the vault docs, PDFs, chat exports, meeting transcripts, screenshots, URLs, the repository you are working in, or the history of past agent sessions. New knowledge merges into existing pages, contradictions are flagged, and a manifest tracks every ingested source so later runs process only what changed. Every claim is tagged as extracted, inferred, or ambiguous, and queries read titles, tags, and summaries before page bodies so their cost stays roughly flat as the vault grows.
Session hooks inject an owner profile, open threads, and recent changes at the start of a session and capture anything reusable at the end. An optional Python CLI and API use the vault as agent memory on disk, with no embedding model, vector store, or network, and the vault can also run as a Docker memory service that agents reach over HTTP or MCP.
On structural questions about a 38-page vault with Claude Sonnet, the project reports these results against an agent without the skills (a small study with two runs per question):
| Measure | Plain agent | With Obsidian Wiki |
|---|---|---|
| Time to answer | 81 s | 19 s |
| Correct answers | 44% | 83% |
| Tool calls | 9.9 | 4.6 |
| API cost | $0.202 | $0.208 |
Features
- Ingest: docs, PDFs, chat exports, transcripts, screenshots, URLs, and the current repository, with code-graph-aware distillation
- Session mining: imports past sessions from Claude, Codex, Copilot, Hermes, OpenClaw, and Pi
- Cited queries: answers carry wikilink citations, plus topic briefings and weekly digests
- Vault hygiene: lint for broken links, orphans, and contradictions, deduplication of synonymous pages, and automatic cross-linking
- Incremental updates: a manifest of ingested sources means later runs process only the delta
- Provenance tags: each claim marked extracted, inferred, or ambiguous, with lint flagging speculative drift
- Session memory: hooks recap the owner profile and open threads at session start and capture reusable knowledge at the end
- Session search: a topic graph over agent session history finds a past session from a vague description
- Graph export: graph JSON, GraphML, Neo4j Cypher, Postgres SQL, or a self-contained interactive HTML graph
- Python API: use a vault as per-user agent memory with remember, search, and recap calls
- Agent support: Claude Code (also as a plugin), Codex, Cursor, Windsurf, Gemini CLI, and a dozen more
