AWS publishes validation pipeline to screen AI agent memory writes

AWS published an implementation that screens Amazon Bedrock AgentCore Memory writes and can allow, quarantine or block content before agents retrieve it in later conversations. The pipeline runs through a Lambda interceptor on AgentCore Gateway, combining checks for semantic divergence, injection patterns, namespace boundaries and unusual write rates. AWS says its demo returned the expected decisions across 11 scenarios, including eight attack payloads. Those tests used a local keyword-overlap approximation rather than the proposed Titan embedding model, so deployed scores and decisions may differ. The gateway validation also does not cover writes made through direct SDK calls that bypass it.