Post

I've been using an LLM memory / wiki tool built around the principles Karpathy shares here where you use the OKF markdown bundle structure as the substrate for a knowledge graph, and then expose very simple query/update/add methods for creating and maintaining that graph.

gist.github.comllm-wikillm-wiki. GitHub Gist: instantly share code, notes, and snippets.

27/7/26, 2:52 pm0 reposts1 like

Josh Hendricks | MVP

@joshooaj.com

27 July

The implementation I've used the most so far is called Understory and it includes a built-in agent "librarian" and an MCP server as the primary external agent interface. Wire up any/all of the agents you use to the MCP and now they all have shared access to the same knowledge graph.

github.comGitHub - thecodacus/understory: understory — memory that grows. Self-wiring, plain-markdown memory for AI agents (MCP + local models + a living graph).understory — memory that grows. Self-wiring, plain-markdown memory for AI agents (MCP + local models + a living graph). - thecodacus/understory

Josh Hendricks | MVP

@joshooaj.com

27 July

But while I've been exploring other implementations of the idea, I've noticed a pattern of content creators sharing prompts instead of software, and suggesting that everyone have their agents build their own agent memory and I'm not sure how I feel about the wastefulness of it.