Xiangchen Cheng, Yunwei Jiang, Jianwen Sun, Zizhen Li, Chuanhao Li, Xiangcheng Cao, Yihao Liu, Fanrui Zhang, Li Jin, Kaipeng Zhang
4 min
As LLM agents are applied to increasingly long-horizon tasks, the standard practice of appending all past observations and tool calls to the prompt leads to context growth, which dilutes attention and obscures the impact of specific memory components. This paper asks whether a bounded, typed memory interface can replace raw transcripts to make agent memory more inspectable, ablatable, and efficient.
The authors propose a "bounded-memory contract" where every decision is made from a fresh user message assembled via typed retrieval. This interface organizes information into five distinct slots: fixed protocol instructions, state-specific schemas, game rules, episodic summaries, and triggered strategic skills. By using Slay the Spire 2—a complex, stochastic, turn-based deck-building game—as a testbed, the researchers evaluate how these specific layers influence agent performance. They implement a fixed-difficulty (A0) ablation study to isolate the impact of each memory layer and provide a public archive of 298 trajectories to support further research into long-horizon agent design.
The study demonstrates that the bounded-memory contract allows for granular control over what information an agent sees at each step. In the fixed-A0 ablation, the baseline agent (no memory or skills) achieved a win rate of 3/10. The addition of a triggered strategic skill layer (L5) increased this to 6/10. While these results are directional, they highlight the potential for typed memory layers to improve decision-making in complex environments. Furthermore, the authors show that this architecture remains operational across different model backbones and supports scaling to higher difficulty levels (A6–A8) when episodic and skill stores are allowed to update.
This work shifts the focus of agent development from simply managing context windows to designing explicit, modular memory architectures. By providing a reproducible testbed and a methodology for ablating specific memory layers, the authors offer a framework for researchers to systematically study how different types of information—such as episodic memory versus strategic skills—shape the long-term behavior of LLM agents.
Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of any single memory component is hard to isolate. We introduce and instrument an alternative bounded contract: every decision is made from a fresh user message assembled by typed retrieval, with no raw cross-decision transcript appended. The prompt thus stays bounded across runs of any length, and any single layer can be ablated in isolation. We instantiate the contract in Slay the Spire 2, a closed-rule stochastic deck-building game whose runs require hundreds of tactical and strategic decisions. A public online benchmark of frontier LLMs on the same game reports zero wins at the lowest difficulty across five configurations, and the developer-reported human win rate at the same difficulty is 16%; the task is hard but not saturated. Within our harness, a fixed-A0 ablation shows the largest observed difference when triggered strategic skills are enabled: the no-store baseline wins 3/10 games and adding the skill layer 6/10. At this sample size the comparison is directional rather than statistically decisive (Fisher exact p\approx0.37); a cross-backbone probe and public accumulating-context baselines are reported as operational comparisons rather than controlled tests of the contract variable itself. We release a reproducible testbed: 298 completed trajectories with condition tags, frozen memory/skill snapshots, prompt records, and analysis scripts -- an agent design and a validated, reusable methodology for studying how explicit memory layers shape long-horizon LLM-agent decisions.
Sam: So the game is designed to expose exactly the kind of flexible thinking the briefing folder is supposed to support.
Alex: That's the logic. And they found that enabling just one of those five layers—the tactical skills folder—produced a noticeable improvement in the AI's performance. That's a meaningful signal, because it tells you something specific: stored, reusable strategies are doing real work, not just adding noise.
Sam: How confident should we be in that finding?
Alex: Cautiously optimistic. The sample size is small, so the researchers themselves describe these results as directional—a promising early signal rather than a settled conclusion. More testing across different tasks and environments would be needed before drawing firm conclusions.
Sam: You mentioned they also used something called "ablation" to test the layers. What does that mean?
Alex: It's a standard technique in AI research. The idea is simple: you test a system by removing one component at a time, then checking whether performance drops. It's like testing a car by pulling out one part at a time to see which ones are actually load-bearing. If you remove the skill library and the win rate falls, you've learned something concrete about what that layer contributes.
Sam: So instead of just knowing that the whole system works, you know why it works—which part is carrying the weight.
Alex: Precisely. That's what makes this approach useful beyond this one paper. It gives researchers a reusable method for studying how different memory layers affect an AI's behavior—not just in games, but potentially in any long-horizon task.
Sam: So the contribution here isn't just a better AI for one game. It's a cleaner way to think about and test AI memory design in general.
Alex: That's the core takeaway. The goal isn't to build bigger agents with more data. It's to build more principled ones—agents that know what to remember, what to set aside, and why. The briefing folder, not the archive.
Sam: That's a meaningful shift in how you'd approach the problem.
Alex: It is. And if that principle holds up under further testing, it could quietly reshape how long-running AI systems are designed. Thanks for listening to ResearchPod.