Image: d2908q01vomqb2.cloudfront.net · rights & removal

Building a context-aware AI assistant on AgentCore and OpenClaw

Reporting by AWS Machine Learning BlogRead the original at aws.amazon.com

Executive Summary

A system is presented for building context-aware AI assistants by combining AgentCore, OpenClaw, and Amazon Bedrock AgentCore memory to provide long-term conversational context. The architecture allows agents to accumulate knowledge through event creation and structured extraction strategies, enabling the assistant to remember past details across sessions. Personalization is achieved by storing user-specific memories in isolated namespaces tied to a chat ID. The process involves retrieving long-term records based on a current query, assembling them with explicit preferences and inferred facts, and injecting this context into the system prompt for model responses. This architecture routes tasks, using Claude Haiku for fast text and Claude Sonnet for vision tasks, and incorporates cost-saving measures like prompt caching to manage inference costs.

Facts Only

* The system uses AgentCore runtime and OpenClaw as an agentic system.
* AgentCore memory turns chats into durable knowledge using structured metadata.
* Input flows from Telegram webhooks and scheduled jobs through API Gateway, Lambda functions, and the InvokeAgentRuntime API.
* Amazon S3 stores workspace, AWS KMS handles encryption, and AWS Secrets Manager holds bot tokens.
* Memory is stored in user-specific namespaces like sprout/{chatid}/longterm for isolation.
* Long-term memory extraction uses strategies like USERPREFERENCE, SEMANTIC, and SUMMARIZATION.
* Retrieval calls RetrieveMemoryRecords against the relevant namespace, capped at 50 results within a three-second budget.
* Text turns route to Claude Haiku 4.5, and vision turns route to Claude Sonnet 4.5.
* Prompt caching is used to reduce inference costs and latency by up to 90% and 85%.
* Skills are manifested via a community-skills.json manifest, making the system domain-agnostic.

Full Take

The pattern demonstrated is the externalization of state management from the stateless nature of large language models into a structured, auditable memory layer managed by an orchestration framework. The efficacy hinges not just on the retrieval mechanism but on the deliberate structure applied to memory—specifically creating clear namespaces and using indexed metadata for filtering—which enforces cognitive boundaries between different user contexts. The cost-saving mechanism via prompt caching highlights a critical tension in agentic design: balancing the need for rich, personalized context with the necessity of keeping inference costs low. Furthermore, routing models based on task (Haiku vs. Sonnet) is an example of architectural segmentation that acknowledges varying computational needs, suggesting that optimizing performance requires matching tool/model selection to the complexity of the required reasoning, rather than applying a single monolithic approach. The design guidelines suggest that true personalization emerges from systematic structure rather than emergent behavior; this implies that building reliable, personalized systems relies on disciplined pre-definition of memory boundaries and prompt formatting before allowing the model to perform its core task.

From the original · AWS Machine Learning Blog

Artificial Intelligence Building a context-aware AI assistant on AgentCore and OpenClaw Off-the-shelf AI assistants answer individual questions well, but they fall short on a different axis: continuity.

Read the full story at aws.amazon.com

Sentinel — provisional

Raw score, 0 to 1 (uncalibrated)0.15

No strong signs of machine writing were found in the source article. Provisional estimate, not a finding that a person wrote it.

The model's own note (unverified)

This text reads like a detailed technical whitepaper or high-level blog post written by an expert detailing the architecture and philosophy behind building context-aware AI agents, demonstrating strong human authorial intent.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

What the model noticed in the source article

Moderate sentence length variance; consistent with technical explanation structure.

High internal coherence; the argument flows logically from problem to solution to implementation details.

Heavy use of specific, technical terminology and precise system architecture description suggests deep domain knowledge by an author familiar with AWS/LLM patterns.

No overt signs of fabricated statistics or highly polished, generic argumentation; the technical specifics anchor the text.

Signs the model read as against machine writing

Specific referencing of proprietary AWS services (AgentCore, Bedrock AgentCore, CloudFormation, S3, KMS) and detailed architectural flow strongly suggest authorship by someone with direct experience or deep domain knowledge.

The nuanced discussion on asynchronous memory extraction, prompt caching costs/latency trade-offs, and namespace design reveals an applied, rather than purely theoretical, understanding of agentic systems.