You know that feeling: you come across a great article online and bookmark it;

you read a valuable tweet and screenshot it;

you encounter an important concept and jot it down in your notes.

And — nothing ever comes of it.

Fragmented information keeps piling up, but very little of it ever becomes truly useful. What we lack isn't in— it’s a system that lets information grow on its own .

What Traditional Note-Taking Can't Do

Most people take notes in one of these ways:

Put simply, these approaches are all about storing , not building . Knowledge needs to be compiled, linked, and continuously maintained before it becomes a real knowledge base.

The Core Idea of llm-wiki: Digest Once, Accumulate Forever

The system's logic is straightforward: You provide a source (a link, an article, a file), the AI digests it into structured knowledge and stores it in an interconnected wiki. When new knowledge comes in next time, it doesn't start from a blank — it merges into the existing knowledge network.

Think of it as a knowledge version of — each new article is a new box of bricks. The AI doesn't just put the whole box on the shelf; it opens it up and snaps the useful pieces onto the structures already standing.

The overall flow is simple:

The next time you ask about the same topic, the AI doesn't re-read the original articles it directly queries the compiled wiki pages. That's "compile once, reuse forever."

How to Use It?

Step 1: Initialize

Create a blank knowledge base with a topic, direction, and an empty homepage.

Step 2: Add Sources

Add sources to your knowledge base — send text, images, PDF files, etc., to the LLM. The model parses them directly or calls vision tools to process files.

Step 3: Digest

The full digestion process consists of two steps:

— Structural Analysis : After reading the source, the AI extracts:

— Page Generation : Based on the analysis, the AI updates the wiki:

The key point: if a concept already has a page, the new content appends to it rather than overwriting it. This way, perspectives on the same concept from different sources naturally converge.

What Makes It Smarter Than RAG?

RAG (Retrieval-Augmented Generation) is what many AI Q&A tools use today — you ask a question, the AI searches the original text, reads it on the spot, and answers. The upside is convenience; the downside is that it reads from scratch every single time — no memory .

llm-wiki works more like a diligent reader taking careful notes:

The key is "compile once" — like writing code. Source code that gets interpreted every time (RAG) is naturally slower; compiled into binaries (wiki), lookups become lightning fast.

A Few Neat Details

Cache Mechanism

You don't feed the same article in twice. llm-wiki has a cache that checks before processing: have I handled this file before? Has the content changed?

Health Check (Lint)

Over time, a knowledge base can develop issues:

The AI can run periodic health checks, listing issues with suggestions for fixes.

Who Is This For?

A Final Thought

We live in an age of information overload. If all we do with the endless stream of content is bookmark, forward, and screenshot it without thinking, it will never become knowledge.

What llm-wiki does isn't complicated — give the AI a container you control, and let it assemble the fragments into a map. Don't measure yourself by the number of notes you have. Measure yourself by whether your knowledge base is growing and connecting.

The best part: all files are local markdown, plain text, ready to open with Obsidian, VS Code, or even Notepad. Your knowledge truly belongs to you, with no dependence on any cloud service.

And of course, this system itself is open source (MIT license). You can find the source code here: https://clawhub.ai

Build your own knowledge base with existing agent tools.