AIToday
Large Language ModelsAI Coding AssistantsAmazon AI BlogPublished: Oct 7, 2026, 06:00 JST

AWS shows Sprout, a $5–9/month AgentCore + OpenClaw assistant

AWS shows Sprout, a $5–9/month AgentCore + OpenClaw assistant

3 Key Points

  1. What happened

    AWS demonstrated Sprout, a gardening assistant built on OpenClaw and Amazon Bedrock AgentCore. It runs on one CloudFormation template, uses Claude Haiku 4.5 for text and Claude Sonnet 4.5 for vision, and costs about $5–9/month as of July 2026.

  2. Why it matters

    AgentCore memory turns one-off chats into durable, structured records, so the assistant recalls preferences across sessions.

  3. What to watch

    The pattern is domain-agnostic; swapping the skills manifest lets the same pipeline serve a support bot or help desk. Long-term memory is extracted asynchronously, so new facts are typically not recalled within the same session.

WHO IT HITSDevelopers and small teams building personal or customer-facing assistants can adapt this pattern without standing up a custom vector store or always-on server. Teams testing it should plan for asynchronous memory extraction and set a budget cap, since costs scale with usage.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The post frames a common frustration with off-the-shelf assistants: they answer well but start every conversation from zero, forcing users to re-explain context such as a garden's soil or a preference for organic fertilizer. AWS's answer is to pair OpenClaw, an open source agentic system, with AgentCore memory, which stores conversation turns as events and asynchronously extracts longer-lived records. Three extraction strategies are configured: USER_PREFERENCE, SEMANTIC, and SUMMARIZATION.

Several engineering choices are load-bearing. Records are filed into per-user namespaces, with the Telegram chat ID as the only variable segment, so two gardeners never mix. A retrieval step runs on every turn, and an assembly function ranks explicit preferences ahead of inferred facts before injection into the system prompt. AWS also routes image turns directly from server.py to the Bedrock Converse API rather than through the OpenClaw gateway, because the in-container build dropped image content parts. Model choice is treated as configuration, not code.

The cost story rests on consumption-based pricing and prompt caching: the stable persona and memory block go first, the volatile user message last, letting Bedrock skip recompute on the unchanged prefix. Whether the pattern holds up for a given team likely hinges on how much memory injection inflates each prompt and how tolerant users are of the extraction delay, since new facts typically become retrievable only in a later session.

FAQ
How much does running this assistant cost?
Light personal use runs about $5–9/month as of July 2026, roughly $2 for infrastructure, $1–3 for Haiku text, and $2 for Sonnet vision. A built-in AWS Budget alerts at 80 percent and 100 percent of a cap you set.
Which AI models does it use?
It routes text chat to Claude Haiku 4.5 and image understanding to Claude Sonnet 4.5 on Bedrock. Model IDs are environment variables, so you can swap models per deployment without rebuilding the image.
How do users talk to the assistant?
Through a Telegram bot. Messages arrive via Amazon API Gateway and a webhook Lambda function, while scheduled jobs such as morning watering reminders arrive through Amazon EventBridge Scheduler.
Amazon AI BlogRead Original Article

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleEmbeddingGemma 2: 740M-param open model goes multimodal