Researchers developed AgenticSTS, which replaces AI agents' growing chat logs with five structured memory layers. Tested on the card game Slay the Spire 2, the agent won 6 out of 10 games, outperforming competing agents that won none. This approach keeps prompts at around 5,000 tokens instead of ballooning past 500,000.
SynthePulse Insight · AI deep reading
Version 2 · 1 source
Traditional AI agents rely on ever-growing chat logs, causing context bloat and inefficiency. AgenticSTS uses five-layer structured memory to double win rates in the card game Slay the Spire 2 while drastically reducing token consumption.
When large language model agents perform tasks involving multi-step decisions, they typically append each step's observation, reasoning, and action to the prompt, causing the context window to expand rapidly. This phenomenon, known as 'context rot,' not only significantly increases token consumption and latency but also can dilute the model's attention and degrade decision quality.
The researchers chose Slay the Spire 2 as the test environment. The game involves hundreds of decisions (card selection, combat, route planning, etc.), has high randomness, and human players achieve only about 16% win rate at high difficulty. Previously, frontier models performed extremely poorly on this game, with no wins in the AGI-Eval evaluation.
AgenticSTS abandons the traditional cumulative log model. At each decision step, it rebuilds the prompt from five fixed memory layers: L1 fixed protocol instructions, L2 current game state and legal action pattern, L3 retrievable game rules, L4 summary of past runs, and L5 trigger-based skill library.
This design ensures a constant prompt length (about 5,000 tokens) and independent management of each layer, making it easy for researchers to pinpoint which component contributes to performance improvements.
At the lowest difficulty A0, the agent with no memory layers wins 3 out of 10 games (30%); after enabling the L5 skill library, the win rate jumps to 60% (6/10). However, researchers note that with only 10 games per condition, this result may be statistical noise.
In difficulty escalation tests, the agent with cross-run memory reaches A6-A8 levels, while the memoryless one stalls at A2-A4. This shows that memory is crucial for handling higher difficulties.
In terms of efficiency, AgenticSTS's actual user text per step is about 5,000 tokens, while the two comparison agents, STS2MCP and CharTyr, due to their cumulative logs, reach up to 527,000 tokens per step call, taking four times as long as AgenticSTS, with 96% of the delay from model response waiting.
Researchers attempted to freeze the memory stack accumulated by Gemini 3.1 Pro and transfer it to Qwen 3.6-27B and Deepseek V4-Pro. The former's average score increased by 84.5%, while the latter's dropped by 18.1%, and neither won a game. This indicates that memory content is strongly tied to the model that generated it, making cross-model transfer unstable.
The main limitations of the study include: only one character (Silent) tested, a total of 50 runs, and no direct comparison with cumulative context methods within the same code framework. The researchers acknowledge that the gap with public agents reflects the current technical ecosystem rather than an ablation effect of a single memory factor.
AgenticSTS demonstrates the great potential of structured memory in solving context bloat: achieving higher win rates with more efficient token usage. However, memory generality, cross-model transfer, and more rigorous ablation experiments remain to be explored. The researchers have released complete run data for 298 games, memory snapshots, and evaluation scripts for further academic study.
This direction, together with projects such as Anthropic's Memory Tool, China's GAM framework, and the open-source Mastra, drives innovation in AI agent memory mechanisms. However, designing memory architectures that are both efficient and general remains an open problem.
Source is the secondary media outlet THE DECODER, reporting based on a preprint paper; the experiment scale is small (50 runs), and the authors acknowledge that conclusions may be affected by noise.
Structured memory is an effective direction for solving AI agent context bloat, but issues such as memory generality and cross-model transfer still require further research.
Primary report
Primary source