Learned memory boosts GPT-4.1 to τ²-bench top without frontier model

altk-evolve, a customer service agent, reached 67.5% pass^4 on τ²-bench retail using gpt-4.1, a non-frontier model. it achieved this by learning guidelines through analysis of training task attempts, which were then added to the agent's prompt. the agent outperformed all retail entries on the october 12, 2026 leaderboard snapshot. the approach uses a basic react agent with a two-phase think → act loop and gpt-5.5 as the customer-simulator model. the learned guidelines were extracted using gpt-4.1 and appended to the system prompt, allowing the agent to follow domain policies and execute correct actions.