Slay the Spire 2 finally cracked: not by bigger models, but by killing the chat log

For months, the smartest AI models on the planet couldn’t win a single game of Slay the Spire 2. The digital card game, where a successful run demands hundreds of sequential decisions, exposed a fundamental flaw not in reasoning, but in short-term memory. Every move an AI made was added to a growing transcript, eventually drowning the model in its own history. The new AgenticSTS project throws that entire playbook out. Instead of a ballooning chat log, the agent rebuilds its context from scratch for every decision using five discrete memory slots—ranging from fixed game rules to a dynamic skill library. The result is immediate. At the game’s lowest difficulty, the agent’s win rate doubled from 30 to 60 percent once the skill library was activated.

The architecture’s real sting is in efficiency, not just wins. Compared to two publicly available agents that use the standard accumulating-context approach, AgenticSTS used 66 to 90 times fewer tokens for every point scored. One competing agent bloated a single model call to over 527,000 tokens by the end of a game. The structured agent keeps the actual prompt steady at around 5,000 tokens regardless of run length. That token bloat translates directly into a fourfold time penalty for traditional agents, with model latency eating up 96 percent of the delay.

But the research also reveals how brittle this learned memory can be. When the team froze a memory stack built by Gemini 3.1 Pro and handed it to other models, the results were a mess. Qwen 3.6-27B saw its score jump 84.5 percent—yet still couldn’t win a game. Deepseek V4-Pro’s performance dropped 18.1 percent. The memory isn’t a universal plug-and-play upgrade; it’s deeply coupled to the model that created it. The authors are upfront about the study’s limits: with only 50 total runs and a single character tested, the doubling of the win rate could still be noise, and the head-to-head comparisons aren’t clean ablation studies since the competing agents also differ in routing and decision logic.

Still, the direction is compelling because it challenges the brute-force assumption that bigger context windows will solve everything. It turns out that what an agent chooses to forget—or rather, how it structures what it remembers—matters far more than how much it can technically cram into a prompt. The next real test will be running an accumulating-context agent inside the same codebase to isolate the memory effect completely.