An international research team has introduced AgenticSTS, a new memory system for AI agents, tested in the card game Slay the Spire 2. Five separate memory layers replace the previously common approach of ever-growing chat logs. In testing, the agents’ win rate rises from three to six out of ten runs, while token usage per decision stays constant.
Five Layers Replace Growing Chat Logs
Conventional AI agents append the entire prior conversation log to the next prompt with every new decision. In long runs, this context grows to roughly 500,000 tokens, according to the team, a multiple of what a single model can still process reliably. AgenticSTS replaces this approach with a bounded contract of five typed layers that are freshly assembled for each individual decision, instead of dragging along raw data from previous turns.
The code published on GitHub distinguishes role templates for different game phases and state schemas with the respective legal actions. Added to this are static game knowledge about cards and enemies as well as a library of strategic rules with clear trigger conditions. The episodic memory layer additionally organizes past runs by character, ascension level, act, and enemy type, so that matching experiences can be retrieved specifically. Each of the five layers can be switched off individually, making its contribution separately measurable.
The testbed is the card game Slay the Spire 2, a rule-based deckbuilder with random elements that, according to the authors, requires hundreds of tactical and strategic decisions per run. The agent controls the game via a programming interface provided by a purpose-built game mod.
Win Rate Rises From Three to Six Victories
In the research group’s tests, the baseline version without the additional skill layer wins three out of ten runs. When the fifth layer with strategic rules is activated, the rate rises to six out of ten runs. According to the AGI-Eval benchmark cited in the paper, frontier language models currently score zero wins at the lowest difficulty level, with a maximum possible defeat threshold of 33 turns. Human players, according to the data, win about sixteen percent of roughly 240 million evaluated community runs. The study thereby addresses an already documented weakness of current language models in very long, strategic task sequences.
With only ten runs per configuration, the authors themselves acknowledge that the difference is not statistically confirmed. A Fisher’s exact test reportedly yields a p-value of about 0.37, meaning the improvement counts as a trend rather than a proven effect. The win rates have not been independently verified so far; they come exclusively from the team’s own experiments. More robust conclusions would require significantly larger samples, as the publication itself points out.
Token Usage Stays Constant Over Long Runs
The central efficiency gain lies less in playing strength than in resource consumption. Classic context-appending approaches grow to up to 500,000 tokens within a single run, while the prompt size in AgenticSTS stays nearly the same throughout the entire run. Compared to open-source competing systems, the approach reportedly requires between 66 and 90 times fewer fresh language-model tokens per point scored, according to the team.
The project is provided by the research lab AlayaLab, which maintains the source code openly on GitHub. For verifiability, the team additionally publishes 298 complete game trajectories with condition tags, frozen memory and skill snapshots, as well as complete prompt records and analysis scripts. On the project page at Hugging Face, the paper, published in early July, ranked second on the platform’s paper leaderboard on its release day. Other research groups can use the provided data and code to test individual memory layers in their own experiments. For development teams, flat token usage potentially means lower compute costs for agents that must make many consecutive decisions.
What will matter is whether the principle of bounded, typed memory layers also holds up outside of game testbeds, for instance in customer-service or coding agents whose context windows grow similarly uncontrolled in practice. It also remains open whether the win rate actually stays stable with larger samples or turns out to be statistical noise. Since the code and raw data are freely available, this question should be settled in the coming months through independent follow-up tests.
