Hasty Briefsbeta

Bilingual

Letting Claude Play Text Adventures

17 hours ago
  • Attended an AI hackathon focused on cognitive architectures (Soar, ACT-R) for scaffolding LLMs.
  • Chose text adventures (specifically Anchorhead) as a long-horizon task to evaluate agent performance.
  • Built a Python wrapper around the dfrotz Z-machine interpreter to control the game.
  • Implemented a 'trivial' harness that keeps full game history, but it becomes too token-expensive.
  • Tested a memory harness (recent 5 turns + semantic memory via tool use), which reduced tokens but caused inefficient wandering.
  • Attempted smaller generated games (e.g., escape-the-room, heist) but found them too linear and uninteresting.
  • Identified future improvements: domain-specific memories, automatic/manual geography tracking, and episodic memory summaries.