How I Built a Persistent AI Persona That Passed Cognitive Testing (And What Broke Along the Way)

The article is loaded on HackerNoon with the meta description at 159/160 characters. The cracked mirror image is showing as a featured image. You need a TL;DR now. Here's one:TL;DR:"Built a persistent AI persona on Claude using externalized Notion memory, 29 voice rules, and a 4-tier loading system. The system scored 413/430 on cognitive assessment, connected test questions across sections unprompted, and used the user's name without being told to. Along the way I discovered the Pocket Watch Problem (context degradation at three scales), fought deference collapse, crashed the context window with memory fetches, and learned that identity might be a spectrum, not a binary. Full architecture docs at veracalloway.com.