ALTK-Evolve uses learned memory to top τ²-bench with a GPT-4.1 customer-service agent

ALTK-Evolve uses gpt-4.1 to analyze agent trajectories on τ²-bench training tasks and distill successes and failures into short behavioral rules, such as authenticating users before order actions, doing one thing per step, and itemizing each refund. The rules are appended as a markdown list after domain policy, with no changes to model weights or the ReAct scaffold. Retail keeps 292 guidelines, lifting mean accuracy from 78.1% to 82.5% and Pass^4 from 57.5% to 67.5%, above all retail entries in the October 12, 2026 leaderboard snapshot. Guideline count must be tuned per domain: retail benefits from more, while in airline 14 high-support guidelines raise Pass^4 to 45% versus 40% for all 101. Cost falls mainly on input, with retail rising from $0.084 to $0.139 per conversation, but assistant turns for the same tasks drop by about 0.3–0.4, cutting end-to-end latency; the guideline block is identical per call, so production can amortize it with prompt caching. In airline, using gpt-5.5 for both agent and guideline distillation raised accuracy from 72.5% to 82.5% and Pass^4 from 55.0% to 70.0%, showing guidelines still help frontier models. Teams building customer-service agents that need higher policy adherence without retraining can reuse this prompt-injection approach directly.