Richmond Alake of Oracle outlines an “intelligence plane” for agents at AI Engineer, with a vendor demo and product overview.

The layers

  • A control plane manages identity, permissions, policies, orchestration and audit.
  • Alake’s proposed intelligence plane spans what an agent knows, remembers, can see and can improve over time.
  • He describes memory as retaining, reusing, recalling and refining information, including working, episodic, semantic and procedural forms.
  • An ontology describes entities and relationships; a semantic layer gives business terms consistent meaning.

The demonstrations

  • A recorded finance-analyst comparison adds an ontology, semantic layer, access controls and caching to a baseline agent.
  • Alake reports 99% accuracy and 32,000 tokens for the enhanced system versus 200,000 tokens for the baseline. The talk does not establish an independent benchmark or isolate which change caused the difference.
  • A second example routes conversation or workflow memories with a fine-tuned small decision model; embeddings and reranking are other proposed components.

Build and govern the memory

  • Capture useful traces, define entities and terms, and make access permissions enforceable at the data layer.
  • Alake describes Oracle products for agent memory, inherited user permissions and private inference; these are vendor-presented capabilities.
  • Fine-tuning on stored traces needs careful evaluation and data governance. The presentation does not prove that its learning loop improves production outcomes.

“Information does not equal intelligence.” — Richmond Alake