Releases·Tool Release
NVIDIA NeMo Agent Toolkit adds Amazon S3 Vectors as custom memory provider on Amazon EKS
An AWS Storage Blog post walks through implementing Amazon S3 Vectors as a custom persistent memory provider for NVIDIA's NeMo Agent Toolkit (NAT), deployed on Amazon EKS using a multi-agent investment research use case. It covers NAT's memory plugin interface, the auto_memory_agent workflow, multi-agent memory sharing via metadata filters, and operational deployment patterns. Practitioners get a concrete recipe for production agent memory with elastic vector storage, strong write consistency, and IAM-based isolation.
0 engagement1 sourceThu, Oct 1, 2026, 05:34 PM
AI Overview
- AWS publishes an implementation guide for building a custom MemoryEditor plugin that uses Amazon S3 Vectors as NAT's persistent memory layer on Amazon EKS
- NAT's memory subsystem exposes MemoryEditor (add_items, search, remove_items), MemoryItem, and MemoryBaseConfig, with built-in providers Mem0, MemMachine, Redis, and Zep
- The plugin generates embeddings with Amazon Titan Text Embeddings V2 into a 1024-dimension S3 Vectors index with filterable metadata and strong write consistency
- S3 Vectors scales to 2 billion vectors per index with configurable cosine or euclidean search and per-tenant indexes secured by IAM policies
- NAT's auto_memory_agent workflow captures user messages and agent responses and injects relevant context automatically without LLM-invoked memory tools
- A three-agent investment research example shares one S3 Vectors index, scoping memory by agent_id, user_id, and team_id metadata filters
- Agents deploy as Kubernetes containerized services via nat serve with IAM access through IRSA and HorizontalPodAutoscaler CPU-based autoscaling; Amazon Bedrock AgentCore is a managed alternative
- NAT's nat eval harness and hyperparameter optimizer tune top_k and similarity threshold, though the post reports directional expectations rather than benchmarks
Entities
Related
Releases·Tool Release
AWS benchmarks OpenSearch Serverless configs for Bedrock Knowledge Bases RAG vector store selection
An AWS Machine Learning blog post guides builders choosing a customer-managed vector store for Amazon Bedrock Knowledge Bases, comparing Amazon OpenSearch Service, Amazon Aurora PostgreSQL with pgvector, and Amazon S3 Vectors across three RAG use cases. It includes original benchmarks on OpenSearch Serverless Classic collections using Amazon's ESCI shopping dataset, measuring NDCG@10, latency, and index size across embedding dimension and data type configurations. The guidance helps practitioners trade off latency, cost, and search quality when building RAG on AWS.
An AWS Machine Learning blog post guides builders choosing a customer-managed vector store for Amazon Bedrock Knowledge Bases, comparing Amazon OpenSearch Service, Amazon Aurora PostgreSQL with pgvector, and Amazon S3 Vectors across three RAG use cases. It includes original benchmarks on OpenSearch Serverless Classic collections using Amazon's ESCI shopping dataset, measuring NDCG@10, latency, and index size across embedding dimension and data type configurations. The guidance helps practitioners trade off latency, cost, and search quality when building RAG on AWS.
0 engagement1 sourcerssThu, Sep 17, 2026, 03:53 PM
Social·Community
Developer shares agent-as-memory-controller architecture using Neon Postgres
A developer describes their long-term memory architecture for AI agents, using an agent-as-memory-controller pattern with hierarchical directory memories stored in Neon Postgres. The approach uses tools for CRUD operations and leverages Neon's scale-to-zero and instant branching for debugging.
A developer describes their long-term memory architecture for AI agents, using an agent-as-memory-controller pattern with hierarchical directory memories stored in Neon Postgres. The approach uses tools for CRUD operations and leverages Neon's scale-to-zero and instant branching for debugging.
8 engagement1 sourceredditTue, Jul 21, 2026, 06:03 PM
Social·Community
AWS demonstrates Kiro and EKS MCP Server workflows for simplifying Amazon EKS cluster management
Medium guide walks through managing Amazon EKS with Kiro and EKS MCP Server
Medium guide walks through managing Amazon EKS with Kiro and EKS MCP Server
0 engagement1 sourcerssSun, Oct 4, 2026, 12:16 PM
Products·launched
Maximem Synap — persistent memory SDK for AI agents
Maximem Synap is a memory SDK that gives AI agents persistent memory across sessions. It uses LLM-based typed extraction and entity resolution to store user preferences and context so agents stop re-asking known information. It targets developers building agents who lose context between sessions.
Maximem Synap is a memory SDK that gives AI agents persistent memory across sessions. It uses LLM-based typed extraction and entity resolution to store user preferences and context so agents stop re-asking known information. It targets developers building agents who lose context between sessions.
37 engagement1 sourceredditSat, Sep 19, 2026, 01:44 PM
Releases·Tool Release
Build a multi-agent music production pipeline on Amazon Bedrock AgentCore Runtime Instances
An AWS post walks through deploying a three-agent music production pipeline on Amazon Bedrock AgentCore Runtime Instances, the new EC2-based compute option for long-running agent workflows. Three Strands Agents (composition, delivery, compliance) colocate on one GPU instance via a shared session ID, render a track with ACE-Step, master it, and screen it against a back catalog. The architecture demonstrates patterns AWS says extend to 3D rendering, simulation, inference, and media processing workloads.
An AWS post walks through deploying a three-agent music production pipeline on Amazon Bedrock AgentCore Runtime Instances, the new EC2-based compute option for long-running agent workflows. Three Strands Agents (composition, delivery, compliance) colocate on one GPU instance via a shared session ID, render a track with ACE-Step, master it, and screen it against a back catalog. The architecture demonstrates patterns AWS says extend to 3D rendering, simulation, inference, and media processing workloads.
0 engagement1 sourcerssWed, Sep 30, 2026, 03:21 PM