A home for personal context

Duncan Davidson's O'Reilly Radar piece (this page is a pointer post summarising it) opens with a simple observation: every agent he uses is building a model of him, and each one keeps that model inside its vendor's walls. Claude has learned his prose; ChatGPT remembers what he's working on; switching products means starting over, and using three agents means three redundant reconstructions of the same person.

His proposal: a canonical, user-controlled repository of personal context that any agent could request permission to use. He frames it as the private counterpart to his earlier essay on personal websites in the AI age — a website is where you teach the world who you are, but the richer private context (preferences, projects, history) currently lives in fragments inside whichever agent you happen to be using.

He situates the idea relative to the Markdown-pile pattern people have already adopted, citing Karpathy's LLM Wiki as the best-known example. A year of working that way taught him five things a personal context system must get right, and raised the deeper question — not which disk the context lives on, but inside which trust boundary. The full argument lives in the Radar article; the post ends with the hope that "changing agents won't mean changing homes."