Amazon unveils ARISE self-evolving agent memory
Amazon researchers introduced ARISE, a self-evolving memory architecture that enables enterprise AI agents to learn from execution failures without retraining model weights or altering workflow logic. By recording failure attributions as retrievable micro-memories and consolidating them into validated macro-rules, the system boosts task accuracy by up to 10.9 percentage points while cutting token overhead.
Fine-tuning foundation models or manually patching prompts every time an enterprise agent fails is unscalable in production. ARISE offers a practical alternative by treating runtime failures as self-healing memory updates governed by strict validation gates.
- –**Two-tier memory structure:** Captures task-specific failure attributions as retrievable micro-memories, then consolidates repeated patterns into compact, generalizable macro-rules.
- –**Strict regression safeguards:** New rules are committed only after passing held-out accuracy checks, preventing policy drift and hallucinated edge-case patches.
- –**Self vs. teacher attribution:** Deploying Claude Opus 4.6 as an offline attribution teacher yielded +10.9 pp on SOP-Bench, while self-attribution using the base model still secured a 10.0 pp gain.
- –**Cost and inference efficiency:** ARISE-Teacher cut total model invocations by 4.0% and output tokens by 5.2%, demonstrating that smarter context reduces agent thrashing.
DISCOVERED
3h ago
2026-11-19
PUBLISHED
4h ago
2026-11-19
RELEVANCE
AUTHOR
arkyyang