Most knowledge workflows re-run retrieval or paste raw logs every time you ask; obsidian-wiki flips that model by treating an LLM as an active maintainer that compiles once and serves many queries. The practical effect is a growing, navigable markdown graph that reduces repeated token cost, preserves provenance, and makes prior work discoverable across projects and agents.
What Sets It Apart
- Skill-based agent integration: All automation is encoded as readable markdown "skills" the agent executes. That means a single repo can be symlinked into Claude Code, Cursor, Codex, Gemini CLI, Pi and other agents so the same instructions run everywhere.
- Compile-once, query-many: Instead of RAG at query time, sources (PDFs, chat JSONL, transcripts, screenshots) are ingested, distilled, merged into canonical pages, and tracked in a manifest so only deltas are reprocessed.
- Practical vault tooling: Built-in features include automated cross-linking, tag taxonomy normalization, graph colorization, provenance flags on claims, export formats (GraphML/Neo4j/HTML), and linting to find orphans or contradictions.
- Multi-agent provenance & targeted ingest: Special skills mine individual agent histories (Claude, Codex, Hermes, Pi, OpenClaw etc.) and allow cross-agent topic-first ingest (e.g.,
/wiki-codex "rust ownership").
Who it's for and trade-offs
Great fit if you regularly reuse insights across codebases or agents and want a single, versioned markdown graph that an LLM maintains for you. It helps engineers who want automated distillation of meeting notes, research papers, chat logs, and project learnings into searchable, linked pages. Look elsewhere if you need a hosted, collaborative knowledge product with real-time multi-user editing out of the box—obsidian-wiki assumes a local Obsidian vault (file-based) that you manage (optionally backed to GitHub) and relies on agent execution to maintain the content. Expect some setup (symlinks/config) per agent and the usual caution when ingesting sensitive personal data into automated pipelines.
