ARISE: Practical Self-Evolving Memory for Accurate and Cost-Efficient Enterprise Agents
Summary
- Researchers proposed ARISE (Agent Refinement through Iterative Self-Evaluation), a memory architecture for enterprise LLM agents that evolves from failures without altering model weights, tools, or workflow logic.
- ARISE uses either the model itself or a teacher model to attribute failures, store corrections as retrievable micro memory, and consolidate them into macro rules verified by held-out accuracy.
- Evaluated across 14 enterprise-workflow domains using benchmarks like SOP-Bench and tau^2-bench, ARISE-Teacher improved accuracy over baselines by 10.9 pp and 5.8 pp respectively.
- ARISE-Teacher reduced model calls by 4.0% and output tokens by 5.2% while achieving comparable per-task costs (-1.0%).
Why it matters
The ARISE architecture offers a cost-efficient, weight-free method to continuously improve enterprise LLM agents via self-evolving memory, directly boosting reliability in dynamic operational environments.