aws-samples Architecture Knowledge
AgentCore Runtime
- Managed container runtime for AI agents
- Supports Strands SDK, LangChain, and custom frameworks
- Request-driven model (invoke_agent_runtime API)
- Session-based with automatic scaling (including scale-to-zero)
- Memory integration (STM events + LTM strategy extraction)
Memory Architecture (Plato Design)
- Platform files (SOUL.md, IDENTITY.md) → baked into container image
- User memory → AgentCore Memory with namespace isolation
- STM: Conversation events per session (list_events / create_event)
- LTM: Extracted records from 4 strategies (Semantic, UserPreference, Summary, Episodic)
- Namespace isolation:
/actors/{actorId}/for server-side filtering - Explicit memory: save_memory / recall_memory tools for high-value facts
Slack Integration Pattern
- API Gateway → Lambda (3s ack) → SQS → Worker Lambda (15min) → AgentCore
- Session ID: thread-based for shared context, channel+user for non-thread
- Actor ID: Slack user ID for memory isolation