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

Learned Memory Propels a Simple ReAct Agent to the Top of τ²‑bench

A recent experiment shows that a modest prompt‑level “learned memory” can lift a standard ReAct‑style customer‑service agent—running on the older GPT‑4.1 model—to the highest scores on the τ²‑bench leaderboard. By extracting short, high‑confidence behavioral rules from the agent’s own training runs and appending them to the system prompt, the team achieved substantial gains in both capability (mean accuracy) and reliability (Pass⁴), all without fine‑tuning or human annotation.

---

From Raw ReAct to Guided Agent

The baseline agent follows the stock τ²‑bench ReAct scaffold: a THINK → ACT loop that uses GPT‑4.1 for reasoning and GPT‑5.5 to simulate customers. The only modification is a block of learned guidelines added after the domain policy. These guidelines are concise markdown bullet points such as:

• Authenticate the user before any order‑related action.
• Confirm all intended changes with a structured summary before execution.
• Take one tool call per turn and wait for the result.

Guidelines are generated automatically: the same GPT‑4.1 model analyzes successful and failed trajectories on the training split, distills recurring patterns, and records the support count (how many distinct tasks a rule appeared in). No human labeling or model weight updates are involved.

---

How Much Guidance Is Right?

Three guideline configurations were evaluated:

| Variant | Support Threshold | # Guidelines (Retail) | # Guidelines (Airline) | |---------|-------------------|-----------------------|------------------------| | Baseline| — | 0 | 0 | | sup3 | ≥ 3 | 31 | 14 | | lossless| ≥ 1 | 292 | 101 |

Retail domain benefitted from the full lossless set, climbing from 57.5 % Pass⁴ (baseline) to 67.5 % (+10 pts) and raising mean accuracy from 78.1 % to 82.5 % (+4.4 pts). The airline domain showed a different pattern: the compact sup3 core gave the best reliability (Pass⁴ = 45 %, +10 pts) while lossless improved accuracy but slightly hurt Pass⁴.

The takeaway is that the optimal amount of guidance is domain‑specific; practitioners should tune the support threshold rather than hard‑coding a single value.

---

Cost vs. Benefit: Tokens, Turns, and Latency

Guidelines increase the input token count because the system prompt grows, but they do not lengthen conversations. In fact, matched‑run analysis shows a modest reduction in assistant turns (≈ 0.3–0.4 fewer turns per solved task). Token‑cost impact per conversation:

• Retail: baseline $0.084 → lossless $0.139 (+65 %).
• Airline: baseline $0.081 → lossless $0.107 (+32 %).

Because the guideline block is static, it can be cached across calls, mitigating the extra cost in production. The net trade‑off is favorable: a modest increase in prompt cost yields higher accuracy, higher reliability, and slightly lower latency.

---

Do Frontier Models Still Need Guidance?

The same learned‑memory pipeline was applied to a GPT‑5.5 agent on the airline split. Even with a stronger model, appending automatically mined guidelines (support ≥ 2) lifted accuracy from 72.5 % to 82.5 % (+10 pts) and Pass⁴ from 55 % to 70 % (+15 pts). Thus, learned memory is not merely a patch for weaker models; it provides a systematic boost even for frontier LLMs.

---

Practical Takeaways

1. Prompt‑level learned guidelines can be generated automatically from an agent’s own training runs, requiring no external data or fine‑tuning.
2. Domain‑specific tuning of guideline support thresholds maximizes the reliability‑accuracy trade‑off.
3. Cost impact is limited to input tokens, which can be mitigated with prompt caching; conversation length actually shrinks.
4. Even frontier models benefit, making learned memory a broadly applicable technique for policy‑bound customer‑service agents.

By turning a simple ReAct scaffold into a “memory‑augmented” system, teams can achieve leaderboard‑leading performance on τ²‑bench without the engineering overhead of model updates—an attractive path for enterprises seeking reliable, policy‑compliant automation.