Everyone is talking about GraphRAG and graph memory for LLMs:

• GraphRAG: an LLM extracts entities and relationships into an explicit knowledge graph.
• Graph memory: facts are edges in a knowledge graph that also remembers when they were true.
• Wiki memory: Markdown pages are the nodes, links are the edges. Karpathy suggests Obsidian’s graph view to see its shape.
• Ontologies: the schema of the graph—which kinds of things exist and how they are allowed to relate.

My post about putting the agent’s memory in the repository is a very, very tiny, limited version of the same idea. Funny how things that have been around for a long time keep coming back.

LLMs are good at plain text files. LLMs are good at Markdown. Markdown can contain links. So of course we write memory into Markdown files, and of course we link those files to each other.

It’s hypertext. It’s the wiki. It’s the World Wide Web… all over again.