What should AI assistants remember permanently?

Crescitaly · reddit · 2026-08-17

The post explores the core challenge of persistent memory for AI assistants: it's not just about remembering more, but deciding what survives, handling conflicts, and dealing with "stale truths" where correct stored facts become wrong later. The author proposes a system separating explicitly confirmed preferences, dated project facts, expiring temporary context, and sensitive details. The OP asks users which control matters most: editable memory, expiration dates, source links, confidence labels, or a reliable forget function.

Original post →

More from AGI Musings

AGI Musings channel →