ARISE lets enterprise agents learn from failures without retraining
A research paper describes a memory architecture where an agent's own failures drive corrections stored as retrievable memory, with no changes to model weights, tools or workflow logic.
In plain words
Large language model agents often repeat mistakes because their underlying model stays fixed. ARISE is a proposed memory system in which the agent analyses its own failures, stores fixes as small retrievable notes, and merges them into broader rules. The model itself is never retrained, so the approach aims to raise accuracy while keeping costs flat.
What was announced
The paper, posted to arXiv, introduces ARISE (Agent Refinement through Iterative Self-Evaluation). When the agent fails, the same model or a separate teacher model attributes the failure and proposes a correction. Corrections are kept as retrievable micro memory and consolidated into macro rules, and a rule update is accepted only if accuracy on held-out data rises. The authors say this lets an agent's memory evolve from its own failures without updating model weights, tools or workflow logic.
In tests across 14 enterprise-workflow domains drawn from SOP-Bench and tau2-bench using five models, the teacher variant, which uses Claude Opus 4.6 for attribution and consolidation, gained 10.9 percentage points over the baseline on SOP-Bench and 5.8 points on tau2-bench. It also outperformed GEPA by 4.5 points on tau2-bench and 1.4 points on SOP-Bench. When the deployment model served as its own attributor, the self variant still gained 10.0 points on SOP-Bench. On cost, the teacher variant used 4.0% fewer model calls and 5.2% fewer output tokens at comparable per-task cost, reported as minus 1.0%.
What it does not show
- The evidence is the paper's abstract only; full experimental setup, baselines and caveats are not available in the supplied passage.
- Results come from the authors' own benchmarks and models; no independent confirmation is provided.
- The abstract does not state whether ARISE is deployed or available in any product.
For security and IT teams
- Scope
- The reported gains apply to enterprise-workflow benchmark domains from SOP-Bench and tau2-bench, not to a named production deployment.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.