Learned memory beats frontier models on customer service benchmark

ALTK-Evolve, a new method detailed in a Hugging Face blog post, demonstrates how learned memory can top the τ²-bench leaderboard for customer service tasks without updating model weights. Using GPT-4.1 as both the agent and the distillation engine, the system analyzes training trajectories to extract behavioral guidelines, which are then appended to the system prompt. This approach increased Pass^4 scores by 10 percentage points in retail and airline domains compared to a baseline ReAct agent. The method relies on prompt caching to manage input costs and reduces conversation length, offering a low-cost alternative to fine-tuning for structured customer service workflows.