“Knowledge graph” sounds like an enterprise data project. The idea underneath is simple, and if you’ve ever linked two notes in Obsidian you already use it. Here’s what a knowledge graph is, how a personal knowledge graph differs from a company’s, and why an AI app answers better with one than with a pile of files.
What a knowledge graph is
A knowledge graph has two ingredients. Nodes are the things you know something about: a person, a project, a contract. Edges are the connections between them. It gets useful once edges have a type, so two nodes aren’t just “somehow related” but connected by “works on”, “is part of” or “is stated in”.
The node “Kitchen renovation” links to the contractor’s quote via “is stated in”, to the contractor via “carried out by”, and to a decision note via “decided in”. None of it is new information. It’s just connected.
A document sits in one folder; a node can hang off any number of edges. And a tag only says “belongs to category X”, while an edge says how two specific things relate.
The famous example: Google
Google introduced its Knowledge Graph in 2012 under the slogan “things, not strings”. Search was meant to know things (people, places, companies) and how they relate, not just match character sequences. The info panel next to the results for, say, an actor comes from that graph.
Enterprise graph vs. personal knowledge graph
Inside a company, the term usually means a smaller version of Google’s project: products, customers, suppliers, modeled in a graph database with a fixed schema and a team to maintain it. That pays off when many systems query the same data.
A personal knowledge graph works the other way round. There’s no schema up front; the structure grows out of your notes. In Obsidian, the best-known tool for this, type [[Kitchen renovation]] in a note and you get a link, and the target note shows it as a backlink. The graph view draws the whole thing as a web of dots. A few hundred notes linked like this are a knowledge graph, whatever you call it. The idea overlaps with the second brain.
The weak spot is manual work: every edge exists only because you set it. And anything that isn’t typed text, a PDF, a scan, a forwarded email, usually stays an isolated dot with no connections.
Why an AI app answers better with a graph
Hand an AI app a pile of documents and it answers from the passages most similar to your question. That works when the answer sits in one place, less well when it’s spread out: the deadline is in the contract, the cancellation in an email, the reasoning in a note from last year.
A graph helps because it knows what’s next to what. Once search has found the contract, the AI can follow the edges and pull in whatever is attached: the email, the note, the person. In 2024 Microsoft Research presented GraphRAG, an approach along these lines: first build a graph of entities and relationships from the documents, then draw on it when answering. It’s aimed mainly at questions that need an overview across many documents, less at looking up a single fact.
The AI app also needs a way into the graph. That’s what MCP is for: the open standard Claude, ChatGPT and other AI apps use to read external data sources.
Building a graph without manual tagging
Linking no longer has to be your job. Language models are good at spotting the people, projects and topics in a text and drawing edges from them. You can keep writing wikilinks and have a tool suggest missing ones, or drop material in and let the graph build itself.
Memexa is one example of the second route (our own product, to be clear). Notes, documents and scans (via OCR) are read and linked automatically to projects, people and topics using typed edges rather than tags. An Obsidian vault imports as a ZIP with its wikilinks kept as edges. Claude, ChatGPT or any other AI app with MCP reads your memory at https://brain.memexa.ai/mcp, and answers come with citations down to the page. Hosted in the EU, exportable as Markdown any time. Here’s how it works in detail. Plans start at €2.49 a month billed yearly or €2.99 monthly (pricing), with no free plan; the 14-day trial requires a card.
If you’d rather keep everything on your own machine, Obsidian plus a local MCP server gets you a long way too. The options compared.
How to start a personal knowledge graph
- Start with what you have: PDFs, existing notes, an email folder. A graph built from existing material is useful on day one; an empty one never is.
- Don’t design a schema. The kinds of nodes you need become clear after a few weeks.
- Link while you write, not afterwards. Typing
[[Name]]takes a second; cleaning up later takes evenings. - Test it with questions that span several documents. “What did we decide about the renovation, and why?” shows quickly whether the connections hold.
- Check the exit: a graph that grows for years should export to an open format like Markdown.
Frequently asked questions
Do I need a graph database?
Not for a personal graph. A folder of Markdown files with wikilinks already is one. Graph databases like Neo4j earn their keep with large datasets and complex queries, not with your notes.