# Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle

## Executive summary

The talk outlines the architectural evolution of AI agents, arguing that the next critical layer after the Control Plane (governance and orchestration) is the Intelligence Plane. This plane moves beyond simply having access to data, focusing instead on how an agent understands, remembers, and learns. Key components include implementing Ontology and Semantic layers, treating memory as structured data, and utilizing decision models for efficient memory routing and continual learning. The speaker emphasizes that owning and structuring memory is the key differentiator for mission-critical AI systems.

## Key takeaways

- Intelligence vs. Control: Having the same information (Control Plane) does not equate to intelligence. Intelligence requires unique interpretation, structured knowledge, and memory (Intelligence Plane). The historical example of WWII codebreakers finding 'AF is Midway' illustrates this difference.
- The Intelligence Plane Components: The Intelligence Plane is an AI infrastructure layer that allows an agent to improve over time. It encapsulates what the agent should know (Knowledge), what it should remember (Memory), what it is allowed to see (Governance/Data Access), and how it gets better (Continual Learning).
- Efficiency Gains via Intelligence: Agents built with an Intelligence Plane (using Ontology and Semantic layers) demonstrated significantly improved accuracy (99% of the time) while drastically reducing token consumption (32K tokens vs. 200K tokens for a naive agent).
- Own Your Memory: The core actionable advice is to treat memory as structured data, capture memory traces, and build systems that enable continual learning. This structured memory is presented as the system's 'moat.'

## Technical details

- Agent Memory Taxonomy: Agent memory is defined by the 'Four Rs': Retain, Reuse, Recall, and Refine. The taxonomy includes short-term, long-term, and cognitive functions (semantic, episodic, procedural, working memory).
- Architectural Layers: The Intelligence Plane introduces critical layers: the Ontology layer (adds accuracy beyond simple keyword search) and the Semantic layer (improves understanding of meaning, e.g., correctly defining 'concentration').
- Security and Data Access: Deep Data Security ensures that the agent inherits the identity and data access permissions of the user it is acting on behalf of, propagating user-level row access into the data layer.
- Memory Routing: Memory routing uses decision models to efficiently and cheaply route output from the context layer or LLM to the correct memory store (e.g., conversation memory or workflow memory).
- Private Inference: For regulated industries (like finance), private AI services allow hosting embedding, LLM, and re-rankers on-premise or in air-gapped infrastructure, ensuring data sovereignty.

## Practical implications

- Architects should shift focus from merely orchestrating agents (Control Plane) to building systems that manage structured knowledge and memory (Intelligence Plane).
- Implement a dedicated Ontology and Semantic layer to move beyond keyword search and achieve higher accuracy in mission-critical workloads.
- Treat memory capture as a primary data source for fine-tuning and continual learning, rather than a separate feature.
- Utilize decision models to automate and optimize memory routing, ensuring context is sent to the correct storage location efficiently.

## Topics

AI Agent Architecture, Agent Memory, Control Plane vs. Intelligence Plane, Semantic Layer, Deep Data Security, Continual Learning, Oracle AI Database, Oracle Agent Memory Package, AIE CODE summit

Source: https://www.youtube.com/watch?v=1_lX0m64-aI