Team:
Aydar Akhmetzyanov (team lead), Lily Zhu, Dhawal Modi, Vince r
Code:
https://github.com/AydarAkhmetzyanov/PackMind-submission/blob/main/src/werewolf_agents/simple_updated_with_memory_and_cot/agent/simple_updated_with_memory.py
Overall Strategy:
SimpleUpdatedMemoryAgent is designed for strategic play in Werewolf/Mafia games, with an intelligent memory and a CoT (Chain of Thought) verification process for enhanced role-based responses. Built on the OpenAI API, this agent is ideal for players who enjoy strategic decision-making and stealthy gameplay, adapting its behavior based on the game’s ongoing social interactions.
Key Features:
- Role-Specific Strategy: Each role (villager, seer, doctor, wolf) comes with tailored instructions. These guide the agent on strategic decisions like avoiding unnecessary accusations and shielding itself from suspicion, staying consistent with the role’s objectives.
- Memory System: The agent tracks player interactions, adding contextually relevant notes—accusations, alliances, or defenses—to its memory. This adaptive memory allows it to evolve based on interactions, simulating a human-like understanding of trust and suspicion among players.
- Chain of Thought (CoT) Verification: Before sending any public response, the agent verifies if the response aligns with its role and objectives. CoT prompts it to assess if it reveals too much or strays from team goals, encouraging smarter in-game strategies.
- Voting & Interaction Management: In response to game phases like “voting” or “discussion,” the agent adapts its behavior—providing hints for stealthy moves or enforcing strict responses, such as naming a suspect when prompted.
SimpleUpdatedMemoryAgent offers dynamic play that enhances both role immersion and success rates in Werewolf/Mafia games, making it a powerful tool for players seeking a nuanced and strategic AI teammate.