Build a Memory-Persistent AI Assistant on AWS

Off-the-shelf AI assistants lack continuity, forcing users to re-explain context each session. This post demonstrates how to build Sprout, a gardening assistant, using OpenClaw on Amazon Bedrock AgentCore runtime, with AgentCore memory turning disposable chats into durable, metadata-tagged knowledge. The architecture is domain-agnostic and deploys via a single AWS CloudFormation template. Two entry points, Telegram webhooks and EventBridge schedules, converge on one agent. Consumption-based pricing keeps costs around one to two dollars monthly for light use, versus roughly thirty-five dollars for an always-on EC2 instance.