AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory

Researchers have developed a memory architecture for AI agents that replaces continuous chat logs with five distinct storage layers, preventing prompt sizes from ballooning during complex tasks. By testing this system on the game Slay the Spire 2, the team successfully maintained a stable token count while achieving a 60% win rate where previous models failed. This approach offers a more efficient way for AI to retain relevant information over long sequences without the processing overhead typically caused by expanding context windows.