Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle
Key Takeaway
Moving beyond standard control planes to implement an intelligence plane with dedicated memory routing, semantic layers, and ontology components allows agents to learn continuously while cutting token consumption and error margins.
Highlights
Agent memory relies on four core mechanisms to retain, reuse, recall, and refine information: retention, reuse, recall, and refinement.
Building an intelligence plane into agentic systems reduces token consumption from 200,000 down to 32,000 tokens while achieving 99% accuracy.
Bob Bemmer coined the term software factory at General Electric in 1968 to describe creating programs in a measurable and repeatable way.
During World War II, analysts in Pearl Harbor used a room in a basement to manually record and decode intercepted communications, proving that stored memory outperforms raw centralized data.
Oracle provides an agent memory system packaged as a Python library that automatically generates context cards without requiring manual engineering.
Timeline
Taxonomy of Agent Memory and Control Planes
Agent memory encompasses short-term, long-term, semantic, episodic, procedural, and working memory structures. While control planes handle orchestration, auditing, and guardrails, modern architectures require systems capable of managing software factories where parallel agents execute systematic workflows.
- Agent memory functions through four core actions: retaining, reusing, recalling, and refining information.
- Control planes act as centralized management systems that govern agent tools, permissions, identity, and policy guardrails.
- Bob Bemmer coined the term software factory at General Electric in 1968 to establish repeatable, measurable software creation workflows.
Intelligence Versus Control Through Historical Precedent
Access to identical models and raw information does not guarantee intelligence or unique differentiation. The intelligence plane encapsulates what an agent knows through knowledge, what it remembers through memory, what it can see through governance, and how it improves through continual learning.
- The intelligence plane cuts across the entire agent stack to allow systems to learn from data and memory.
- During World War II, a small group of code breakers stationed in a Pearl Harbor basement successfully decoded that AF stood for Midway by maintaining historical records on paper.
- Centralized organizations with identical information and greater resources can arrive at entirely different interpretations if they lack robust memory practices.
Demonstrating the Intelligence Plane and Semantic Layers
Governed analysts built with an intelligence plane utilize ontology and semantic layers to answer complex domain questions accurately, whereas naive agents relying purely on keyword search fail. These structures optimize cost, accuracy, and latency, known as tokenomics, while incorporating prompt caching and strict guard controls.
- Integrating an intelligence plane reduces token consumption to 32,000 tokens compared to 200,000 tokens for naive agents.
- Ontology and semantic layers decrease the margin of error in mission-critical sectors like finance.
- Decision models, embedding models, and re-rankers operate across retrieval pipelines to optimize production accuracy.
Infrastructure Components and Continual Learning
Engineers must treat memory as data, capture memory traces for fine-tuning, and build systems that enable continual learning across token, weight, and latent spaces. Tools like the private agent factory and open-source memory routers give developers the infrastructure needed to maintain proprietary memory moats and optimize model performance over time.
- Oracle provides a Python package for agent memory that handles memory engineering and automatically generates context cards.
- Deep data security propagates user identity and access permissions directly into the data layer so agents inherit exact authorization boundaries.
- Private AI service containers support on-premise inference for embeddings, large language models, and re-rankers.
Description
Richmond Alake was talking about agent memory before most people had heard of it, and now he argues the next layer after the control plane is an intelligence plane: what an agent knows, what it remembers, what it's allowed to see, and how it gets better. Using the WWII codebreakers who worked out that "AF" meant Midway, he shows why having the same information as everyone else isn't the same as intelligence. He then demos an agent built with an ontology and semantic layer that gets answers right 99% of the time using 32K tokens, versus 200K for a naive agent, and covers Oracle's agent memory package, deep data security, private inference, and fine-tuning small decision models for memory routing. Chapters: 0:00 Intro 0:20 Agent memory in one minute 1:39 The control plane 2:43 The software factory 3:54 The intelligence plane 4:38 AF is Midway 7:37 Where it fits in the agent stack 8:16 What the intelligence plane is 11:33 Agent = model + harness 13:13 Why now: decision models 14:01 Demo: governed vs. naive agent 17:07 What your team should do now 18:07 Agent memory, security and private inference 20:52 Demo: fine-tuning memory routing 24:41 Wrap-up: own your memory Richmond Alake https://www.oracle.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!