NVIDIA · NCP-AAI

The memory and grounding thread

Where an agent's knowledge actually lives. Opens in M1 with memory as an architectural concern, deepens in M5's five-category memory taxonomy and planning strategies, and closes in M6's RAG pipeline, vector retrieval, and GraphRAG/HybridRAG as the mechanisms that ground an agent in knowledge outside its own weights.

NCPA-T2 · 18 lessons across 3 modules

  1. M1M1-01Agent architecture styles: reactive, deliberative, and hybrid systems
  2. M1M1-02Structuring multi-step reasoning: logic trees, prompt chains, and adaptable architecture
  3. M1M1-03ReAct: interleaving reasoning and acting
  4. M1M1-04Multi-agent orchestration topologies: centralized, decentralized, federated, and hierarchical
  5. M1M1-05Memory as an architectural concern: short-term vs. long-term
  6. M1M1-06Knowledge graphs for relational, multi-hop reasoning
  7. M1M1-07Designing the human-agent interface as the oversight surface
  8. M5M5-01Why agents need memory at all
  9. M5M5-02The memory taxonomy: short-term, long-term, episodic, semantic, and procedural
  10. M5M5-03Reasoning frameworks: chain-of-thought and task decomposition
  11. M5M5-04Planning strategies: the five directions
  12. M5M5-05Stateful orchestration: where memory and planning meet
  13. M5M5-06Adapting reasoning from prior experience and feedback
  14. M6M6-01RAG fundamentals: the canonical retrieval pipeline
  15. M6M6-02Vector databases and approximate nearest-neighbor retrieval
  16. M6M6-03Beyond vector RAG: GraphRAG, HybridRAG, and agentic RAG
  17. M6M6-04ETL and data quality for the knowledge an agent retrieves
  18. M6M6-05Real-time reasoning over structured and unstructured knowledge

Part of the throughlines running across the NCP-AAI prep course.