What did AWS publish?
AWS describes a working architecture, rather than announcing a finished business assistant. Its example, Sprout, is a gardening assistant built with OpenClaw, an open source agentic system, running on AgentCore runtime. AgentCore memory stores context that the assistant can use across conversations.
According to AWS, the system accepts Telegram messages and scheduled requests. OpenClaw supplies the agent loop, tools and skills. The deployment sits in a single AWS CloudFormation template. AWS says the same pattern can be adapted for a support bot or an internal help desk by changing the persona and skills manifest.
The key design is a two-layer memory system. Conversation turns become short-term events. Asynchronous extraction then produces long-term preferences, inferred facts and session summaries. On later messages, the assistant retrieves relevant records and adds them to its context. The details appear in the AWS guide.
What could persistent memory mean for an SME?
For a small or medium business, continuity can matter as much as a good individual answer. A support assistant could use earlier context instead of asking someone to explain the same issue again. An internal help desk could retain relevant preferences across conversations. These are possible applications, not business results demonstrated by the gardening example.
The useful distinction is between remembering more and remembering the right information. AWS gives explicit preferences priority over inferred facts. It also describes indexed metadata that can narrow retrieval by topic. A business version should define which records matter for each task rather than loading every past conversation into every reply.
Memory also needs boundaries. Sprout separates records into per-user namespaces, according to AWS. For a business deployment, decide whether information belongs to an individual, a customer account or a team. Then test those boundaries before using real customer or employee data.
What limits and costs should managers check?
AWS says long-term extraction happens asynchronously. A newly mentioned fact may become retrievable in a later session rather than immediately. The example also continues without memory if retrieval fails or times out. That keeps the conversation moving, but a business should decide when missing context requires clarification or human review.
Treat stored context as useful evidence, not guaranteed truth. Inferred facts can be wrong, and preferences can change. Before deployment, establish how people can correct outdated information and how the assistant should handle conflicting records.
AWS gives an estimated AgentCore runtime baseline of approximately $1–2 per month for light personal use, as of July 2026. That is not an SME operating budget. Estimate costs for your own message volume, model usage, memory and supporting services, then measure them during a controlled pilot.
How should a small business take the next step?
Start with one narrow workflow where repeated explanations create extra work. Define what the assistant should remember, what it should never retain and which actions need approval. Use fictional or non-sensitive records to test recall, corrections, user separation and responses when memory is unavailable.
Check the technical requirements before committing. AWS lists Amazon Bedrock AgentCore access, access to the chosen models, a Telegram bot token and basic familiarity with agent orchestration and CloudFormation. Docker and a configured AWS Command Line Interface are needed only when building and pushing your own image. Assign someone responsibility for monitoring, data handling and ongoing maintenance.
If you are also considering digital workers, Terabot gives businesses AI workers with their own computers in private virtual machines. They work in the apps a team already uses and work around the clock. It is in private preview, by invitation only. Assess any option against your workflow, data boundaries and supervision needs.
Key takeaways
- AWS published a build guide for persistent assistant memory, not a ready-made SME service.
- Useful memory requires relevant retrieval, clear user boundaries and a way to correct outdated context.
- Pilot one workflow and measure accuracy and operating costs before expanding.
Written with AI assistance from the source linked above, and checked against it before publishing. Product names belong to their owners. Check the original source before relying on details.