19 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.