The idea: compile once, instead of searching every time
Most AI knowledge bases work with retrieval-augmented generation (RAG): at question time, the system searches the raw documents and hands the best chunks to the model. Karpathy’s point in the gist is that the model is then “rediscovering knowledge from scratch on every question.” His alternative is to let an agent do the work up front. Each new source is read once and folded into a wiki of summaries, entity and concept pages, comparisons and an overview, which he calls “a persistent, compounding artifact.”
His own analogy: “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.” The agent writes; the human reads, browses and asks questions.
The three layers
| Layer | What it is | Who writes it |
|---|---|---|
| Raw sources | Articles, papers, images, data. Never changed. | You collect them |
| The wiki | Linked Markdown pages: summaries, entities, concepts, an overview | The LLM, entirely |
| The instructions | A file such as CLAUDE.md or AGENTS.md describing structure, conventions and workflows | You and the LLM together |
Two files hold it together. index.md lists every page with a link and a one-line summary, and the agent reads it first when answering. log.md is an append-only record of what was ingested and when, with a fixed prefix per entry so it can be searched with simple command-line tools.
The three operations
- Ingest. The agent reads one new source, discusses the main points with you, writes a summary page, and updates the index, the log and every related page. A single source can touch ten to fifteen pages. Karpathy prefers to ingest one source at a time and stay involved.
- Query. The agent searches the wiki and answers with citations to its pages. A good answer can be filed back as a new page, so the wiki grows from questions too.
- Lint. A periodic health check for contradictions, stale claims, orphan pages, missing concepts and missing links.
The gist names supporting tools: the Obsidian Web Clipper for collecting sources, Obsidian’s graph view, the Dataview plugin for queries over page metadata, Marp for slides, and git for version history.
The misunderstanding: it is about Obsidian
Most coverage focuses on the Obsidian setup. The real shift is who maintains the knowledge. Karpathy argues that people abandon wikis because upkeep grows faster than their value, and that an LLM makes upkeep almost free. The format is ordinary Markdown; the method is an agent that reads, links, updates and checks. That also means the wiki is only as good as the model’s judgment: a wrong summary or a wrong link is written with the same confidence as a right one, which is why the lint step and the git history matter.
What this means for a team
The pattern fits a researcher or a founder who curates their own sources. In a team, the open questions start where the gist stops:
- Who edits? One agent owning the wiki works for one person. With five people and their agents ingesting at the same time, edits collide and someone has to review what becomes the shared truth.
- Who may see what? A folder of Markdown has no per-page permissions.
- Which tools read it? Claude Code and Codex can read a Markdown folder in a repository. The colleague in ChatGPT or Copilot cannot, unless someone connects it.
- Where do the lessons go? The wiki compiles sources. Rules a team learns from corrections in daily work need their own path into it.
For the general approach to what such a knowledge base should contain, see How do you build a knowledge base for AI agents? and for the broader picture, What is a knowledge base in AI?
Where Bombos fits
An LLM Wiki works well for one person; Bombos is for the team. It is a plugin: a shared memory layer underneath every AI tool the team uses, from Claude and Codex to ChatGPT, Copilot, Grok and Cursor, that learns how your team works and shows with every answer which notes it rests on. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.
Related questions
Is LLM Wiki a tool I can install?
No. It is an "idea file": a description you paste into an agent such as Claude Code, Codex or OpenCode, which then builds the specific setup with you. Many people have published their own implementations since.
Do I need Obsidian?
No. The wiki is plain Markdown. Karpathy uses Obsidian to read it, follow links and see the graph view, but any Markdown editor or a git repository works.
Does it replace RAG?
At moderate scale it can. Karpathy notes that an index file is enough at around 100 sources and hundreds of pages; beyond that he suggests a search tool such as qmd, a local Markdown search engine with BM25 and vector search.
Can a team use it?
Karpathy lists a business or team wiki fed by Slack, meetings and calls as one use case. The gist leaves out who may edit what, how conflicting edits are reviewed and how other AI tools read it.
