llm-wiki
Your notes, compiled.
A knowledge base your AI agents maintain under written law — not a chatbot, not a vector store.
31 agent skills, a zero-dependency CLI, and an Obsidian vault in the layout you choose, governed by a constitution.
Project site ·
Quickstart ·
The Constitution ·
Skills ·
CLI ·
Report Bug ·
Request Feature
- About The Project
- Getting Started
- Usage
- The Constitution
- The Skills
- Configuration
- Roadmap
- Contributing
- License
- Contact
- Acknowledgments
About The Project
Most "chat with your notes" tools retrieve at question time: you ask, a retriever grabs some
chunks, a model improvises an answer, and nothing is learned. Ask the same question tomorrow and
the same work happens again.
llm-wiki inverts that. Knowledge is compiled once, when a source arrives, into small linked
markdown pages you can read yourself. Your sources are the source code. Your agents are the
compiler. The wiki is the build artifact — and Obsidian is the IDE you browse it in.
The hard part isn't the compiling; it's keeping a compiler honest across many sessions and more
than one agent. So the vault ships with a constitution (AGENTS.md): eleven numbered laws,
a six-point definition of done, and a deterministic gate that decides whether an operation
actually completed — instead of asking the model to grade its own homework.
Laws, not tips. Every rule has a number, a never, or a check.
The project site covers the same ground, renders the guide and the laws, and
draws a link graph from a real vault. Its source is in site/.
How it works
flowchart LR
A["LAYER 1 — SOURCES<br/>PDFs, articles, papers<br/>chat exports, images<br/>immutable ground truth"]
B["LAYER 2 — THE WIKI<br/>small linked pages<br/>frontmatter + provenance<br/>the compiled artifact"]
C["LAYER 3 — THE SCHEMA<br/>AGENTS.md constitution<br/>31 skills + CLI gate<br/>how it stays honest"]
A -->|agents distill| B
C -.->|governs every write| B
B -->|you read and curate| A
Drop a paper in _inbox/, say "ingest my inbox", and an agent distills it into your vault's page
categories — merging into pages that already exist rather than piling up duplicates, marking what
it inferred, updating the index and the op log, and committing. Re-ingesting an unchanged source
is a no-op, because every source is content-hashed in .manifest.json.
The graph weaves itself
Pages are worth more connected than filed. Say "link my pages" and cross-linker scans the vault
and writes the wikilinks that should already exist between related pages. Ask "what connects X and
Y?" and wiki-query walks typed relationship edges, multi-hop, rather than grepping for keywords.
Because the link index is a first-class artifact, graph-query answers structural questions from
that index alone — without reading a single page body — and graph-analyse surfaces god nodes,
communities, and connections you didn't know you'd made.
Compared to RAG
| Typical RAG / notes chatbot | llm-wiki |
|---|
| When work happens | At query time, every time | Once, at ingest — results are reused forever |
| What you get | An answer you can't inspect | Markdown pages you can read, edit, and diff |
| Storage | Vector DB, embeddings, an index | Plain files in a git repo. No database |
| Duplicate handling | Near-duplicate chunks pile up | Law 6: one concept, one page — new info merges in |
| Hallucination | Invisible, mixed into prose | Marked inline: ^[inferred], ^[ambiguous] |
| Trust | "The model said so" | Deterministic gate: doctor + lint must pass |
| Lock-in | Rebuild when the tool dies | It's a folder of markdown. Nothing to migrate |
| API keys | Required | None — your agent already has model access |
Built With
Deliberately boring. The CLI is ~4,400 lines of Python standard library with zero runtime
dependencies, so it cannot break from a transitive upgrade, and the vault is plain markdown
that outlives any of it.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
