NVIDIA and Amazon publish a blueprint for persistent AI-agent memory
AI agents · Friday, 2 October 2026
Why it matters
For an agent founder, the blueprint points to an external vector-storage pattern for long-term memory rather than keeping memory inside a single conversation context. The stated 2-billion-item capacity and pay-for-use model could support large-scale personalization, but the limited testing and unresolved privacy questions make memory governance and evaluation important product requirements.
What happened
NVIDIA and Amazon released a developer blueprint connecting NVIDIA’s NeMo Agent Toolkit with Amazon S3 Vectors so agents can store and retrieve information across separate conversations. S3 Vectors is described as supporting up to 2 billion stored items with usage-based cloud pricing, while NeMo provides the integration layer for agent applications. The setup is not a ready-to-use consumer assistant: it requires technical deployment on AWS, and the article says it has had few real-world tests so far, leaving privacy and access-control questions for developers.
Players & places
- NVIDIA
- Amazon
- AWS