Observed facts: IBM Research reports that GPT-4.1 retail Pass⁴ rose from 57.5% to 67.5% on held-out τ²-bench tasks when the full learned guideline set was appended to the prompt. In an airline test using GPT-5.5, the 103-guideline configuration raised Pass⁴ from 55% to 70%. The authors say the guidelines were distilled from training-task attempts and applied unchanged on held-out tasks.
Mechanism: The method distills behavioral rules from successful and failed training-task trajectories and appends those rules to the agent's system prompt; it does not update model weights.
Why now: The source tests both a non-reasoning GPT-4.1 agent and a GPT-5.5 airline setup, addressing whether learned guidance is useful only as a patch for a weaker model. The reported gains also extend to the GPT-5.5 airline setup.
Implications: The reported gains support treating memory as a policy-bearing interface between past attempts and future actions, rather than merely as a store of retrieved text. Teams can test whether distilled procedural rules improve repeatability without fine-tuning. The article also reports domain differences: a smaller high-support set was more reliable than the full set for GPT-4.1 airline tasks, so guidance volume may need tuning rather than maximizing.
Evidence boundary: These are results reported by the authors on τ²-bench, with 40 held-out retail tasks and 20 airline tasks in the GPT-4.1 experiments, plus 20 airline tasks for GPT-5.5. The benchmark uses simulated customers, written policies, and a strict task outcome; the source does not establish production customer-support impact. The reported benchmark results do not establish real-world production impact.
Watch condition: A replication should test whether the gains persist across customer-service systems, policy sets, and independently evaluated live workflows. Independent tests should check whether the gains persist outside τ²-bench.