What it is
A bounded-memory LLM-agent testbed for Slay the Spire 2 that replaces growing transcripts with typed retrieval, enabling clean memory-layer ablations over long runs. Instead of appending full cross-decision history, each action is made from a fresh prompt built by typed retrieval over fixed rules, episodic summaries, and triggered skills. This keeps context bounded and makes memory layers individually ablatable.
Key results
- Skill-enabled cells won 6/10 at A0 vs 3/10 for no-store baseline
- Prompt-only won 4/10 at A0, below all skill-enabled variants at 6/10
- Released 298 completed trajectories with condition tags, prompts, and snapshots
Shows that bounded, typed memory turns long-horizon agent memory into a reproducible and separately testable evaluation surface.