Most teams I see are teaching their AI agents with Markdown files. `CLAUDE.md`, `AGENTS.md`, architecture notes, prompt snippets, rules in prose. I do it too. It works better than nothing. But it still feels like handing the agent a pile of notes and hoping it reads the right paragraph at the right time.

I spent the last months researching and prototyping a different substrate: a living graph where capabilities, use cases, tickets, decisions, scenarios, runtime facts and code are connected. Not docs about the project. Project knowledge as something queryable.

This talk is a field report from that research. I will show the direction that looks promising to me: a small Flutter app read through a living graph, with use cases opened as playable scenarios that also runs as tests. The question is not "how do we write better Markdown for agents?" The question is what changes when we stop using Markdown as the source of project memory.