Claude Code-based AI builds external memory and 15-min town

The setup described is not a single model but a collection of parts added incrementally in response to concrete failures. Summaries dropped context, so the AI once asked the developer to repeat something already discussed. Notes existed but went unread, so a reminder hook now places relevant records in front of the AI before it answers. On one evening the AI stopped and nobody noticed until midnight; on another day a login expired and went unnoticed for three hours, prompting an alarm and a patrol.

The most unusual part is the small town. Because a conversational AI stops when it has nothing to do, the developer placed a clock and a world outside it, with events that occur independently. The AI checks in when it wakes, and its actions persist as diary entries, delayed letters, and residents' speech. Two memory-related incidents appeared within 5 days of starting: a four-day-old watchdog comment arrived again by town post, and a resident later used a word the AI had said and nearly forgotten. Repetition also appeared, such as a resident saying the same word twice in a day.

Which parts actually help is still not sufficiently verified, according to the author. The piece functions as a map of a trial, not a proven design, and the next installments are to cover where each part stumbled.