FrontierThe story, in brief
AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory
From 500K to 5K tokens. Researchers just solved the agent memory problem that's been tanking LLM performance in long-horizon tasks.
The KeyNews take
Why it matters
Structured memory architecture is emerging as a critical capability gap in agent systems. This research demonstrates a practical solution to token bloat that directly impacts agent reliability and cost efficiency—relevant to anyone building production AI agents.
The key facts
5 to knowAgenticSTS replaces growing chat logs with 5-layer structured memory architecture
Prompt size reduced from 500K+ tokens to ~5K tokens
Agent wins 6/10 games vs 0/10 for competing agents on Slay the Spire 2
Tested on complex long-horizon decision-making game environment
Addresses context window and token efficiency for agentic systems
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: The AgenticSTS project replaces the ever-growing chat log of AI agents with five separate memory layers. Tested on the card game Slay the Spire 2, the prompt stays at around 5,000 tokens instead of ballooning past 500,000. The agent wins 6 out of 10 games, while competing agents don't win any.