Yannic Kilcher: Titans — Learning to Memorize at Test Time (Google Paper Analysis)

Yannic Kilcher·medium signalXBlueskyLinkedInCopy link

Deep paper analysis of Google's Titans architecture, which introduces a neural memory module that learns to memorize information at test time rather than relying solely on in-context attention. The result is a transformer variant that outperforms standard attention on long-context tasks by combining fast in-context retrieval with slower persistent memory, directly relevant to building agents with reliable long-horizon state. Paper originally published January 2025; Yannic's video is the definitive technical breakdown.