Knowledge graphs and the LLM Graph Wiki
Data that's linked and defined before the model shows up is what makes inference trustworthy. Michael Angeles builds knowledge graphs and second-brain vaults, work he trained for in library science.
- 01 What is an LLM graph wiki?
- 02 How do I structure my notes so AI can actually use them?
- 03 What is a knowledge graph and does my company need one?
- 04 Can I hire someone to structure my company's knowledge for AI?
- 05 What does information architecture have to do with AI?
- 06 What's the difference between RAG and a knowledge graph?
- 07 How do you build a second brain that Claude can read?
What is an LLM graph wiki? #
An LLM graph wiki is a body of plain-text notes where every idea is its own atomic page and the links between pages carry meaning, so a language model can traverse it as a graph rather than search it as a pile of documents. Michael Angeles builds these for individuals and teams.
The word “wiki” is doing real work there. A wiki is already a graph: pages as nodes, links as edges. What most wikis lack is discipline about what a page is and what a link means, which is exactly what a model needs to reason instead of guess.
Make every note one idea. Give the relationships names. Define your terms once, in one place, and link to that definition everywhere else. Do that and an agent can answer “what do I know about X and how did I come to think it” by walking edges, rather than by hoping a similarity search surfaces the right paragraph.
Unicron is the working proof. It’s a Mac app that turns a plain-markdown vault into a daily briefing powered by Claude, built on an ontology that manages the graph underneath.
How do I structure my notes so AI can actually use them? #
One idea per note, plain markdown, explicit links between them, and a consistent vocabulary for the things you write about most. Michael Angeles builds personal vaults on this pattern with Claude Code as the engine, so an agent can treat the vault as memory rather than as search results.
Almost everyone starts with the daily-log vault: one long file per day holding six unrelated thoughts, which is fast to write and nearly useless to a model, because retrieving the paragraph you wanted drags along the five you didn’t. I did it for years.
Plain markdown over a proprietary format, always. No lock-in, works with git, and Claude reads it fluently without a converter in the middle. The file system is a perfectly good database when the files are disciplined.
What is a knowledge graph and does my company need one? #
A knowledge graph stores facts as entities and named relationships rather than as rows and columns, so systems can traverse meaning instead of joining tables. You need one when the same concept means different things in different systems and nobody can say authoritatively which is right.
The honest test: ask three teams what a specific number means and see whether you get three answers. If you do, no amount of AI on top of that data will produce trustworthy inference, because the model will confidently average contradictions.
A lot of companies reach for a graph database when what they actually need is a shared vocabulary, which is the half of the work Michael Angeles gets called in for. The database is the easy part.
Can I hire someone to structure my company's knowledge for AI? #
Yes. Michael Angeles takes on knowledge structuring work through Konigi, defining the terms, controlled vocabularies, and concept relationships a team needs before its knowledge graph is worth querying. Scoping is per engagement. This is library and information science applied to AI systems.
For teams, the deliverable usually looks like a defined vocabulary and a concept model: what the organization’s core entities are, what each one means, what relates to what, and which system is authoritative for each. Unglamorous work. It also decides whether the AI layer sitting on top of it can be trusted.
This is work I trained for in an MLS program and have been doing since Bell Labs, well before anyone was putting a model on top of it.
What does information architecture have to do with AI? #
Information architecture is the practice of defining what things are called and how they relate, which is precisely what a model needs in order to reason over your data. Michael Angeles, a co-founder of the Information Architecture Institute, argues AI made the discipline more valuable rather than less.
For twenty years information architecture got sold as navigation design, which undersold it badly. The core of the discipline was always the vocabulary and the relationships between things, and that core turns out to be the exact input a language model needs to reason rather than guess.
Claude will generate a plausible sentence about anything. What it can’t do is know that your “customer” and your “account” are the same object under two names. Somebody has to say so first.
What's the difference between RAG and a knowledge graph? #
RAG retrieves passages that look similar to your question and hands them to a model. A knowledge graph stores explicit, named relationships the model can traverse. RAG is easier to stand up and fails quietly on questions that need several linked facts; a graph handles those but has to be built.
Similarity search is very good at “find me the paragraph about X” and quite bad at “which of our customers in this segment also did that thing, and when.” That second question needs the relationships to actually exist somewhere, and no amount of embedding proximity will infer them for you. This is the distinction a knowledge graph exists to hold.
Most useful systems are both, and that’s how Unicron is built. Structure the parts that carry the relationships, let retrieval handle the prose. Retrieval alone won’t get you to reasoning.
How do you build a second brain that Claude can read? #
Start with a folder of plain markdown files, one idea each, linked to one another, and point Claude Code at it. Michael Angeles built Unicron on exactly this: a local vault, no proprietary format, no lock-in, with the model reading files directly rather than through a sync service.
Keep it local and plain, because that outlives the tool. Vaults built this way work with Obsidian, work with git, and will still open when today’s apps are gone. Unicron is free. It reads a vault it didn’t create, which is the test I’d apply to anything you’re considering.
One habit does most of the work: when you learn something, write the note as if a stranger will read it without the conversation around it. That’s the same discipline that makes a passage citable to an answer engine, which is not a coincidence.
Still have a question?
Tell me what you're building, what a good outcome looks like, and when you'd want to start. That's enough for me to reply with a time, or a straight no if I'm not the right person for it.
Reach out to work with meI reply to every note within two business days.