Building a context-aware AI assistant on AgentCore and OpenClaw

This article demonstrates how to build a context-aware AI assistant called Sprout using Amazon Bedrock AgentCore and OpenClaw that maintains continuity across conversations by remembering user preferences and garden details.

• AgentCore runtime provides consumption-based pricing (~$1-2/month for light personal use) instead of always-on compute costs
• OpenClaw provides the agent loop, tool use, and skills system adapted to AgentCore's HTTP protocol contract
• Two-layer memory system: short-term events store raw transcripts, long-term extraction creates durable structured records with USER_PREFERENCE, SEMANTIC, and SUMMARIZATION strategies
• Per-user namespaces isolate memories so no two chats mix, with indexed metadata keys for filtering by type, section, and plants
• Dual model routing uses Claude Haiku 4.5 for text and Claude Sonnet 4.5 for vision tasks to optimize cost and quality tradeoffs
• Prompt caching reduces inference costs by up to 90% by caching stable persona and memory prefix across requests
• Telegram provides serverless webhook-based front door requiring no client development
• Full system deploys via single CloudFormation template with optional one-command build and deploy script

The architecture generalizes across domains—swap the persona and skills manifest to serve support bots, fitness coaches, or help desks while maintaining the same cost-effective, memory-enhanced agent pattern.