The AgenticSTS project made an AI agent win Slay the Spire 2, a game where frontier models used in classic mode (prompts with ever-growing conversation logs) had never completed a single run. The difference isn’t in the model — all use Gemini 3.1 Pro — but in the architecture: the agent never sees its own chat log, and rebuilds every decision from five separate memory slots (fixed instructions, schema of current state, rules retrieved on demand, summaries of previous runs, library of strategies for recurring situations). Win rate doubles when the strategy library is active, and with persistent memory across runs the agent reaches difficulty A6-A8, where without memory it stops at A2-A4.

Why it matters to you. An agent that accumulates conversation logs grows in cost and latency with every step, until it saturates the context window or dilutes the model’s attention across hundreds of thousands of tokens. AgenticSTS shows that an architecture with external structured memory consumes 90x fewer tokens per point scored compared to traditional agents, and finishes runs in a quarter of the time. It’s the same pattern we use in the course when discussing how to give memory to agents: the log is not the state, and if you treat it as such you pay a price with every decision.

In depth

The problem of the growing log

Classic agents like ReAct or Reflexion append every observation, tool call, and self-reflection to the next prompt. It works for short tasks, but a Slay the Spire 2 run lasts hundreds of decisions (cards to pick, paths on the map, fights to plan). Frontier models tested on AGI-Eval never won a single run across five different configurations. The log grows until it saturates the window or causes the model to lose focus on what actually matters.

The five memory slots

AgenticSTS inverts the pattern. Every decision starts from scratch: no accumulated log, but five independent slots filled on demand.

• L1: fixed protocol instructions, always the same.
• L2: schema of current state with valid actions (cards in hand, enemies, stats).
• L3: game rules retrieved when needed.
• L4: summaries of previous runs, updated between sessions.
• L5: library of strategies for recurring situations (skills), generated or hand-written.

Anything the agent wants to carry from the previous step must first land in one of these slots. The prompt stays short regardless of run length, and each layer can be isolated to understand what actually improves performance.

The numbers

The team tested five configurations across ten runs each at the lowest difficulty level, A0. Without memory, the agent wins 3 out of 10 runs. With the strategy library active (L5), the rate rises to 6 out of 10, whether using hand-written or template-generated skills. Researchers acknowledge that with only ten runs per config, variance could explain the doubling, but the jump is clear.

Episodic memory (L4) doesn’t help at A0, but becomes critical when facing increasing difficulty: with persistent memory the agent reaches A6-A8, without it it stops at A2-A4. The transferability test shows that memory built by Gemini 3.1 Pro doesn’t carry to other models: Qwen 3.6-27B improves by 84.5%, Deepseek V4-Pro drops by 18.1%, and neither ever wins.

The comparison with log-accumulating agents

The most interesting comparison isn’t internal but against two public agents using accumulated logs: STS2MCP and CharTyr. All three agents use Gemini 3.1 Pro for strategic decisions. Both competitors never win in the five runs tested, and for every point scored they send 66 to 90 times more tokens to the model than AgenticSTS uses. In STS2MCP, a single model call late in a run reaches 527,000 tokens because the entire history gets resent with each new decision. AgenticSTS keeps the actual prompt around 5,000 tokens for the entire run.

Accumulation also costs in time: traditional agents take four times longer to reach the same score, and according to provider logs 96% of the difference comes from model latency, not control software.

What remains open

The paper doesn’t say whether the five slots are universal or specific to the video game. Memory built by one model doesn’t transfer to others, limiting reuse. Win rate at the lowest level stays below 60%, and researchers admit ten runs per configuration leaves room for variance. But the pattern — external structured memory instead of accumulated log — is the same you see in real agentic systems and in the memory agent paper, and here you watch it work on a task with hundreds of chained decisions.