Across corporate deployments, ARISE-Teacher, which uses Claude Opus 4.6 for failure attribution and consolidation, delivers a 10.9 percentage point gain over the baseline on SOP-Bench and 5.8 percentage points on τ2-bench. As foundation models iterate rapidly and corporate standard operating procedures shift, enterprise systems face escalating compute bills from ever-expanding context windows. Running agents against changing business policies usually forces engineering teams to constantly rewrite prompts, retrain underlying models, or maintain bloated prompt logs. The operational challenge centers on whether engineering teams can establish automated failure correction without losing control over procedural workflows and budgets.

The mechanism resolves this dilemma through structured memory management rather than continuous model fine-tuning. Amazon researchers Jiangyun Zhang, Zhou Tang, Kristen Surrao, Zhiyu Chen, Albert Guan, Tor Nitayanont, Yupei Zhang, Roopali Singh, Yunkai Xiao, Shreya Pandit, Julia Huang, Shayan Ali Akbar, Erwin Cornejo, Yuan Li, and Yi Zhang developed ARISE, which stands for Agent Refinement through Iterative Self-Evaluation. The framework creates a self-evolving memory loop where an agent refines its behavior based on past operational errors without updating model weights, external tools, or underlying workflow logic. This design holds infrastructure costs steady by preventing prompt expansion on routine tasks while systematically correcting repeated execution mistakes.

The Dual-Memory Validation Loop

How does this framework update execution rules without introducing arbitrary operational drift?

An offline feedback loop intercepts and evaluates failures using historical task traces, agent outputs, and post-task feedback. When an agent fails an assigned run, an attributor module evaluates the trace against the expected outcome or diagnostic data. ARISE uses the model itself or an external teacher model to analyze why the failure occurred and generate concrete, scoped corrections.

"ARISE keeps the corrections as retrievable micro memory, consolidates them into macro rules, and accepts a rule update only if held-out accuracy rises."

By converting raw failure traces into bounded guidance, the system stores individual fixes as micro memory before synthesizing them into macro rules. A held-out validation gate accepts a rule update only when accuracy improves on a dedicated verification split. If a proposed macro rule fails to improve benchmark accuracy or degrades edge-case handling, the gate discards it entirely. This screening prevents unvetted procedural modifications from reaching production runtime environments.

Quantitative Inference Gains

Many operators assume that higher agent precision inevitably inflates inference token consumption and per-query operational expense. In fact, ARISE-Teacher cuts model calls by 4.0% and reduces output tokens by 5.2% while maintaining a comparable per-task cost of -1.0%. ARISE-Teacher also outperforms GEPA by 4.5 percentage points on τ2-bench and 1.4 percentage points on SOP-Bench. Because the runtime agent queries structured macro rules and targeted micro corrections instead of carrying raw conversational histories, the model reaches decisions with fewer execution steps.

Organizations without access to large teacher models can deploy the mechanism using their primary runtime model. With the deployment model acting as its own attributor, ARISE-Self achieves a 10.0 percentage point gain on SOP-Bench. This self-attribution configuration allows teams to eliminate external teacher dependencies while maintaining disciplined procedural adherence across varied enterprise tasks.

To audit current agent spending, review production execution logs from the past month to isolate recurring procedural task failures. Feed those traces through an offline self-attribution gate that tests candidate rules against a held-out validation dataset before deploying new instructions into runtime prompts.

Why it matters

Boosting agent accuracy usually inflates compute bills, yet ARISE-Teacher actually trims output tokens by 5.2%. This divergence occurs because the system turns raw failure traces into compact rules instead of endlessly feeding bloated dialogue history into the prompt. Engineering teams no longer need to spend hours rewriting prompts by hand just to plug recurring execution errors. Slashing procedural errors without altering underlying weights turns prompt bloat into a controlled, automated operational cost.

Our stories: Persistent Memory Boosts R&D AI Agent Success Rates by 90% (August 14); Persistent Memory for UI Agents Cuts Token Burn and Latency (September 8).

Artificial IntelligenceLarge Language ModelsAI AgentsCost ReductionModel Inference