How to give an AI agent semantic memory without a vector database

This article explores how to implement durable, semantic memory for serverless AI agents using Amazon DynamoDB's native vector search and Amazon Bedrock embeddings. By storing vectors directly alongside operational data in DynamoDB, developers can eliminate the need for dedicated vector databases or complex synchronization pipelines. The author demonstrates that this approach is significantly faster than using S3-based vector storage, achieving roughly 2.5x faster writes and 1.4x faster searches in p50 benchmarks. The guide details the three layers of agent memory—state, session, and long-term memory—and provides a practical implementation using the Strands Harness. This method is particularly effective for operational data in high-frequency agent workflows, while S3-based storage remains better suited for large, static datasets like document repositories. The provided demo includes setup scripts, embedding helpers, and performance metrics to help developers integrate semantic memory into their own AWS Lambda-based agent architectures.