My AI Agent Remembered Staging. It Still Didn’t Know If It Was Healthy

说明与用途

实验表明，Bedrock 代理能够跨会话记住用户偏好，但仍需查询当前来源以获取部署健康状态，明确区分记忆与实时证据的边界。

问题

Determining whether an AI agent should rely on remembered preferences or always query a current source for deployment health.

结果

The agent successfully remembered user preferences across sessions, but correctly identified that remembered data does not guarantee current health, enforcing a boundary between memory and real‑time evidence.

工作流

• 1. Provide the agent with a read‑only deployment status tool via AgentCore Harness → AgentCore Gateway → Deployment_status Lambda.
• 2. Add AgentCore Memory with USER_PREFERENCE strategy to store preferred environment.
• 3. In a session, set preference: "My preferred environment is production."
• 4. Query deployment health; agent recalls preference, calls tool, returns fresh result.
• 5. Start a new session; observe delay before memory is available; after processing time, preference is recalled.
• 6. Use a different actor; confirm preference is actor‑specific.
• 7. Change preference to "staging"; observe that tool only supports production and agent correctly reports unknown health.
• 8. Conclude rule: use memory to decide what to look up, but always use a current source for truth.

证据

The Lambda supported one environment: production. It returned a sample service version, instance count, alarms, health status, and a fresh observation time.

来源

Long-term memory is not an immediate key-value write.

来源

The agent remembered that I preferred staging, but it could not retrieve staging health. Instead of guessing, it said that the current health was unknown.

来源