obsidian-llm-wiki

An agent-maintained personal knowledge base. Raw inputs — voice notes, web clippings, typed thoughts — flow in at one end; an LLM agent does all the tedious bookkeeping (reading, summarising, cross-linking, indexing) and a clean, compounding wiki comes out the other end. You curate and direct; the agent files.

It's built on plain Markdown in an Obsidian vault, version-controlled with git, and wired together with a small pipeline of local scripts. The whole "brain" — how the agent thinks about your knowledge — lives in a single human-readable file, CLAUDE.md, that you edit to evolve the system.

No vendor lock-in by design. Everything is Markdown + git + shell. The LLM backend is swappable (a claude -p wrapper, an OpenAI/local-model call, GitHub Copilot CLI — anything that takes a file path and edits files). The repo ships dormant: nothing calls a model until you wire one in.

The idea

Most note-taking dies from maintenance cost. You capture eagerly for a month, then the filing, linking, and tidying outpace the value and the whole thing rots. This project's bet is simple: make maintenance cost ~zero by handing it to an agent, so the knowledge actually compounds.

The design splits the vault into three layers with one hard rule — raw is immutable, refined is the agent's: