llama-wolf

We build a social-deduction game arena (Werewolf) to test a practical memory architecture for language-model agents. Instead of retrieving long chat logs, our agent learns strategy memory: compact, reusable snippets that map situation → tactic → why. After each game, we distill tactics from logs (positive and negative), keyed by role (Villager/Seer/Werewolf), phase (accuse/defend/vote/night), and a small set of cue features (e.g., day bucket, counterclaim status, pressure, claims). At decision time, the agent filters by role/phase, recalls top-k tactics via a vector search over a canonicalized “situation string,” We run a teacher→student setup: GPT-5 and Llama-Maverick generate diverse training games; we extract and score tactics (win rate, vote-shift, acceptance/backfire, recency). At runtime, only Llama plays, with and without memory, against GPT-5 and Llama opponents under identical prompts and sampling to isolate memory effects.