Council

Good vs. bad behavioral examples showing what excellent play looks like vs. common mistakes. Derived from xuyuzhuang-Werewolf contrastive prompting and Avalon-LLM self-play reflection.

Faction-conditional strategies for deception and detection, derived from xuyuzhuang-Werewolf and LLMWereWolf role strategies. Evil agents get deflection and alibi techniques; good agents get detection heuristics.

3-tier memory tracking (STM/Episodic/Semantic) derived from ai-traitors-v2 and ai-town generative agent reflection. Helps agents maintain beliefs about each player and synthesize round memories into strategic conclusions.

Enriches GameMaster narration with awareness of social dynamics, derived from Sotopia 7-dimensional social intelligence evaluation. Tracks influence shifts, trust erosion, and narrative tension cues.

Internal 5-step reasoning pipeline derived from xuyuzhuang-Werewolf. Agent performs Situation-Suspicion-Reflection-Strategy-Response analysis implicitly before generating any output.

Good vs. bad behavioral examples showing what excellent play looks like vs. common mistakes. Derived from xuyuzhuang-Werewolf contrastive prompting and Avalon-LLM self-play reflection.

Faction-conditional strategies for deception and detection, derived from xuyuzhuang-Werewolf and LLMWereWolf role strategies. Evil agents get deflection and alibi techniques; good agents get detection heuristics.

3-tier memory tracking (STM/Episodic/Semantic) derived from ai-traitors-v2 and ai-town generative agent reflection. Helps agents maintain beliefs about each player and synthesize round memories into strategic conclusions.

Enriches GameMaster narration with awareness of social dynamics, derived from Sotopia 7-dimensional social intelligence evaluation. Tracks influence shifts, trust erosion, and narrative tension cues.

Internal 5-step reasoning pipeline derived from xuyuzhuang-Werewolf. Agent performs Situation-Suspicion-Reflection-Strategy-Response analysis implicitly before generating any output.

Talk-frequency awareness, anti-repetition rules, and natural conversation pacing derived from LLMafia dual-module conversation system.

Personal sub-goal framework layered on top of win conditions, derived from Oxyde goal-driven NPCs and ai-town planning system. Adds emotion-goal interaction for more dynamic, motivated behavior.