The researchers picked the deck-building roguelike Slay the Spire 2 as their test bed. A single playthrough involves hundreds of decisions, from picking cards and planning fights to choosing routes on the map and buying items. The rules translate fully into text, randomness is high, and runs are long. Human players win 16 percent of the time on the lowest difficulty level, A0, according to the developers. Frontier models used in the AGI-Eval assessment didn't win a single game across five tested setups. The game is hard but open-ended enough that architectural differences show up clearly.
Typical LLM agents like ReAct or Reflexion append past observations, tool calls, and self-reflections to the next prompt. The context grows with every step until the window overflows or the model's attention gets diluted. AgenticSTS does the opposite. For each decision, the prompt is freshly built from five clearly separated slots.
L1 holds fixed protocol instructions, L2 holds state schemas with currently valid actions, L3 holds retrieved game rules, L4 holds summaries of previous runs, and L5 holds strategy skills triggered for specific situations. Anything the agent wants to carry over from a prior decision must first be written into one of these storage areas. This keeps the prompt short no matter how long a run lasts and because each layer is addressed on its own, researchers can pin down which component actually drives an improvement.
For the main comparison, the team ran five setups against each other, ten runs each at the lowest difficulty, A0. Without any memory layers, the agent wins 3 out of 10 games. Once the L5 skill library is turned on, storing tactical rules for recurring situations, the win rate jumps to 6 out of 10. That holds whether the skills were written by hand or generated from templates.
With just ten runs per condition, the authors themselves admit the doubling could be noise. Episodic memories of past runs (L4) don't help at A0 and only matter when the agent, in a second test mode, attempts the next higher difficulty after each win. With active memory that updates between runs, the agent reaches levels A6 through A8. Without it, runs stall at A2 through A4.
In a separate experiment, the researchers tested whether knowledge built by one model transfers to others. They froze a memory stack that Gemini 3.1 Pro had accumulated during its own games and passed it unchanged to two other models. Qwen 3.6-27B saw its average score rise by 84.5 percent, while Deepseek V4-Pro's score dropped by 18.1 percent. Neither model won a game. The memory contents appear tied to the model that created them and don't transfer well.
More interesting than tweaks to the team's own codebase is the comparison with two publicly available Slay the Spire 2 agents that follow the classic growing-transcript pattern, STS2MCP and CharTyr. All agents used Gemini 3.1 Pro for strategic decisions. Neither competitor won any of its 5 runs.
The cost numbers are even more striking: for every point the two competitors score, they send 66 to 90 times as many tokens to the language model as AgenticSTS does. The reason is the growing log. In STS2MCP, a single model call near the end of a game hit about 527,000 tokens because the entire game history gets resent with every new decision. AgenticSTS keeps the actual user text at around 5,000 tokens no matter how long the game has been running.
The accumulating agents pay a time penalty too, taking four times as long to reach the same level. According to the provider's stats, 96 percent of that time loss comes from model latency, meaning the pure wait for the language model's response rather than the control software around it.
The authors acknowledge this isn't a clean ablation study. STS2MCP and CharTyr also differ from AgenticSTS in routing and decision batching, so the gap reflects the current state of the public landscape rather than the isolated effect of the memory concept, they say.
The team still hasn't run the real test: running an accumulating context within the same codebase using the same scoring. The key metrics are based on 50 runs, and so far only one character (Silent) has been tested on a single game version. Whether the approach works for other characters and patches remains open.
The team is publishing 298 complete game runs, frozen memory snapshots, and evaluation scripts on Hugging Face, so other research groups can test alternative memory architectures in the same environment. The paper's actual claim is more modest than the numbers might suggest when stacked against the competition. By splitting an agent's memory into clearly named layers, you can later figure out which layer drives which behavior.
Efficient memory for AI agents is a busy research area right now. The chat log that grows with every turn makes models slower, more expensive, and less accurate. Researchers call this problem "context rot."
AgenticSTS isn't alone in tackling it. Anthropic uses Memory Tool and Context Editing to automatically strip outdated tool results from the context and store important info in external files. In their own tests, this cut token use for a 100-round web search by 84 percent. The Chinese GAM framework splits archiving and retrieval across two specialized agents. The open-source framework Mastra condenses conversations into concise text notes stored outside the context window, modeling how humans handle memory.