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

ALTK-Evolve is a system that improves AI agent performance by analyzing its past successes and failures on training tasks. It uses a large language model, GPT-4.1, to distill these experiences into a set of short, behavioral guidelines or a "playbook." These learned guidelines are then appended to the agent's system prompt, guiding its future actions in similar situations. This approach enabled a basic ReAct agent with GPT-4.1 to top the τ²-bench leaderboard for customer service tasks, outperforming other entries without any model weight updates.