AWS published an implementation guide for connecting NVIDIA NeMo Agent Toolkit agents to persistent Amazon S3 Vectors memory on Amazon EKS.
AWS published a technical guide on October 1, 2026, showing how developers can connect the NVIDIA NeMo Agent Toolkit (NAT) to Amazon S3 Vectors as persistent memory and run the agents on Amazon Elastic Kubernetes Service (EKS). The post is an implementation walkthrough, not an announcement of a new NAT or S3 Vectors product.
The guide describes NAT’s memory plugin interface and a custom provider that implements its add, search and removal operations. In the example, the provider creates embeddings with Amazon Titan Text Embeddings V2 and stores them in an S3 Vectors index, alongside metadata such as user, team and agent identifiers. The sample index uses 1,024 dimensions to match the embedding model, and the configuration shows NAT automatically saving conversation messages and retrieving relevant memories for responses.
AWS uses a three-agent investment research workflow to illustrate shared memory: a research agent gathers information, an analysis agent looks for patterns, and a synthesis agent prepares reports. Metadata filters let agents retrieve shared team knowledge, while the example also describes periodically consolidating short-lived observations into more durable semantic memories. The agents are packaged as containerized services and deployed on EKS, with access to the vector store scoped through AWS identity and access controls.
The article lists S3 Vectors’ support for indexes of up to 2 billion vectors, metadata filtering and strong write consistency as relevant capabilities. These are service characteristics cited by AWS, not performance results from a comparative test. Its suggested evaluation compares runs with and without memory, but the post explicitly characterizes expected changes in groundedness, token use, latency and duplicate work as directional rather than benchmarked.
AWS also flags data-handling decisions that come with persistent memory: teams should define retention, remove outdated records, avoid putting personally identifiable information in metadata, and restrict access. The walkthrough therefore offers developers an architecture and sample deployment pattern, while leaving the memory policy and workload-specific evaluation to each implementation.