Your notes, ghostwritten by a local LLM that actually remembers

What it does

obsidian-llm-wiki is a Python CLI that ingests Markdown files from a raw/ folder, uses a local LLM (Ollama by default) to extract concepts, and generates interlinked wiki articles inside an Obsidian vault. It compiles incrementally, preserves hand-edits, and can answer questions from your accumulated knowledge without embeddings or a vector database.

The interesting bit

The project implements Andrej Karpathy’s “LLM Wiki” pattern: notes are treated as source material, not final artifacts. The LLM does the bookkeeping — cross-referencing, updating, and linking — while you just keep dumping in raw thoughts. Reject a draft with feedback and the next compile incorporates your critique; five rejections auto-blocks a concept until you intervene. It’s a rare case where “the AI remembers” isn’t marketing fluff because the wiki literally persists as Markdown on disk.

Key highlights

• Fully local-first — Ollama by default, with optional OpenAI-compatible endpoints (Groq, Together, vLLM, etc.)
• Incremental compiles — only articles linked to changed source notes get rebuilt
• File watcher — olw watch auto-processes anything dropped into raw/
• Hand-edit protection — detects manual changes in Obsidian and skips regeneration
• Query without embeddings — olw query answers from the published wiki; --synthesize saves answers as permanent pages
• Self-maintenance — auto-extracts aliases, repairs broken wikilinks, lints orphans, and commits every action with [olw] prefix for olw undo
• Multi-language — auto-detects note language at ingest, no hard-coded word lists

Caveats

• Maintenance mode — new features moved to successor project Synto; this repo gets bug fixes only
• Requires model management — you need to pull and configure separate “fast” (routing) and “heavy” (writing) models; minimal setup works with one model but quality likely suffers
• Obsidian-centric — the wiki structure and wikilinks are designed specifically for Obsidian; other tools may struggle with the conventions

Verdict

Worth a look if you’re an Obsidian user sitting on piles of unstructured Markdown who wants a private, compounding knowledge base without subscription APIs or cloud vector stores. Skip it if you don’t use Obsidian or need actively maintained feature development — the author has already moved on to Synto.