Your ideas are scattered all over the place: in memos, browser tabs, and chat logs that you'll never find again once you close them. Every time you sit down to work, you have to reconstruct context from memory, then start a new chat and explain everything all over again—because AI has no idea who you are, what you're working on, or what you said three days ago.

In April 2026, Andrej Karpathy proposed a solution he calls LLM Wiki: a vault that holds all your knowledge, where AI handles reading, connecting ideas, and growing the collection. It's no longer a chat box that forgets everything after the conversation ends—it's a memory bank that remembers things you've long forgotten yourself.

An open-source project called claude-obsidian, created by AgriciDaniel, has turned this idea into a ready-to-use tool, and it already has over 10,000 stars on GitHub. It turns Claude Code into a librarian for your Obsidian vault. Drop an article, a PDF, or a transcript into it, and it will automatically read, link, and file the content into your existing Markdown knowledge network.


## What Problem Does It Solve

The core logic is simple:

- Drop in articles, PDFs, or transcripts, and it automatically archives everything

- Every note is automatically linked to related existing notes

- Contradictory information is flagged instead of being silently ignored

- All content is plain Markdown stored locally

This setup doesn't lock your content away in a cloud proprietary database, doesn't force you to adopt a specific methodology, and supports PARA, Zettelkasten, LYT, and general-purpose modes. It adapts to your existing way of thinking, rather than forcing you to switch to an entirely new system.

## It's More Than Just a Pile of Notes

Most people organize notes and then forget about them right after saving. claude-obsidian works differently. It follows a repeating cycle: preserve sources, record evidence, connect knowledge, and put it back into use.


Source files and citations are never discarded. Every note links back to traceable original material, and unsupported claims and conflicting content are kept visible. This is the fundamental difference between claude-obsidian and ordinary "just save text" solutions.

In Obsidian, you'll see a truly organic, growing knowledge graph:



## Two Tools, One System

The entire solution only requires two components:

**Obsidian** handles storage. It's a free note-taking app that saves all content as plain text files on your local device. Notes link to each other to form a visible map of your thinking.

**Claude** acts as the brain. It reads your entire vault, places new information where it belongs, links it to existing content, and answers questions that span all your historical notes.

Because all content is plain text, you aren't locked into any specific AI model. If you switch to a different AI next year, you just point it to the same vault and everything keeps working. What you own is the knowledge base itself, not just a tool that sits on top of it.

## Quick Start

The installation process is straightforward:

1. Install Obsidian and create a new vault

2. Download Claude Desktop (claude.com/download)

3. Clone the project and initialize:

```

git clone https://github.com/AgriciDaniel/claude-obsidian.git

cd claude-obsidian

```

4. Initialize your standalone knowledge vault:

```

python3 scripts/claude-obsidian.py init "$HOME/Documents/MyKnowledgeVault" \

--generated-at "$(date -u +%Y-%m-%dT%H:%M:%SZ)" --operation-id "init-reviewed"

```

5. Open the new directory in Obsidian and access it via Claude Code:

```

cd "$HOME/Documents/MyKnowledgeVault"

claude --plugin-dir /absolute/path/to/claude-obsidian

```

Then drop your sources into the `inbox/` folder and run `/claude-obsidian:wiki-ingest`. Drag in a PDF, URL, or meeting transcript, and it will extract people, tools, and companies mentioned, identify core concepts, create or update corresponding wiki pages, link them to your existing content, and flag any points that contradict your existing notes.

## 15 Capabilities, One System

The project comes with 15 built-in capabilities that cover the full knowledge management workflow:

| Capability | Purpose |

| --- | --- |

| `wiki` | Initialize or take over a vault, diagnose readiness status, and assign tasks |

| `save` | Save an answer or insight instead of automatically transcribing it |

| `wiki-ingest` | Turn imported sources into linked pages with source attribution records |

| `wiki-query` | Read-only mode that retrieves evidence from your vault to answer questions |

| `wiki-lint` | Reports dead links, orphaned notes, missing metadata, and outdated indexes |

| `autoresearch` | Bounded web research that explicitly requests external access |

| `wiki-fold` | Creates traceable summaries of operation logs |

| `wiki-mode` | Supports four organization methods: General, LYT, PARA, Zettelkasten |

| `wiki-retrieve` | Context prefixing, BM25, with optional cosine re-ranking |

## Trust Is Built Into the Architecture

One notable feature of this project is how it handles write operations. It doesn't allow multiple agents to write randomly at the same time. Parallel workers only return drafts and evidence, which are checked by an orchestrator before being applied as a single atomic, recoverable transaction. Every knowledge operation includes pre-checks, SHA-256 hashing, and an operation ID, and conflicts are flagged instead of being silently overwritten.

Permission control is also clearly defined: if an agent can technically delete files or send emails, you have to assume it will eventually do so. That's why the project uses read-only and scope-limited keys, rather than relying on prompt constraints alone.

The project is released under the MIT license, designed following Andrej Karpathy's LLM Wiki pattern, and uses kepano/obsidian-skills as a reference base for Obsidian Markdown and base structures.

## Don't Let Your Knowledge Vault Become Just Another Folder

It's common for people to say "I organized it beautifully, then never opened it again"—that's a common flaw of all knowledge management systems. claude-obsidian solves this by using scheduled tasks to let the vault maintain itself. You can set up recurring tasks via Claude Desktop's Schedule feature, so when you wake up each day, your knowledge vault is already automatically organized.

Another key component is the `rulings.md` file. Every time you correct Claude, you add a line to this file, and it won't make the same mistake again. `CLAUDE.md` defines what the system is, while `rulings.md` records the rules it has learned—they are kept separate.

At its core, you aren't just setting up a Claude configuration—you're building your own memory system. It gets smarter every day from the content you feed it. Most people who see a tutorial won't take action. Those who do will open their knowledge graph a month later and see things they had long forgotten, and they'll never go back to a blank chat window again.

Project repository: github.com/AgriciDaniel/claude-obsidian