ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model

ALTK-Evolve uses GPT-4.1 to learn and apply behavioral guidelines for customer service tasks, achieving a 67.5% success rate on the τ²-bench retail benchmark without updating model weights. This approach distills task successes and failures into prompt-based rules, outperforming other retail entries on the leaderboard.