{ "video": { "id": "1_lX0m64-aI", "title": "Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle", "duration": 1547, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "1_lX0m64-aI", "title": "Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle", "one_liner": "Everyone has the same models and information, so agents only stand out through an 'intelligence plane': a layer running through the whole agent stack that covers knowledge (ontology and semantic layers), memory, governed data access and continual learning. Alake's key advice is to own your memory and capture memory traces as fine-tuning data.", "summary": "Richmond Alake (Oracle) argues that information does not equal intelligence. He uses Station Hypo, the Pearl Harbor codebreakers who worked out that 'AF is Midway' from the same intercepts Washington had, to show that memory and interpretation beat centralized control. He traces the field from prompt layers to context layers to today's control plane (identity, policy, orchestration, auditing), and proposes an intelligence plane that cuts across the whole agent stack so agents get better over time. His demos are a governed analyst against a naive one, and his open-source 'Memories' tool, which fine-tunes small decision models for memory routing. He then maps the reference architecture to Oracle products.", "key_points": [ "Agent memory is the mechanism an agent uses to retain, reuse, recall and refine information (the 'four Rs'). Its taxonomy runs short-term/long-term, then semantic, episodic, procedural and working memory.", "The control plane is centralized governance (identity, policy and guardrails, orchestration and permissions, auditing). The intelligence plane cuts through the whole stack and covers 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).", "'Software factory' is not new to the agentic era: Bob Bemer at GE coined it in a 1968 paper about producing programs in a measurable, repeatable way. Alake likens the intelligence plane to the 'knowledge plane' proposed in networking on top of the control and management planes.", "His equations: Agent = model + harness; control = governance + security; adding knowledge, memory and learning gives a 'multiplier effect'.", "Demo, governed analyst vs naive analyst: the governed agent, built on ontology and semantic layers, gets answers right 99% of the time and uses about 32,000 tokens against the naive agent's 200,000, with guardrails, a cache and prompt caching. Without the semantic layer the naive agent got 'what does concentration mean?' wrong. The ontology improved the 'counterparty' answer over plain keyword search.", "A gap of around 5% matters in finance and other mission-critical work; in fraud handling it can mean 'a 100 million fine' or not.", "Second demo, the open-source 'Memories' tool: a Qwen model ('equine' in the auto-transcript) fine-tuned into a decision model for memory routing, sending output to conversation memory or workflow memory. Its training data came from his own agent conversations, which he labelled by memory type. He calls memory routing, embeddings and re-ranking the lowest-hanging fruit for fine-tuning, with quality gains and sometimes lower latency.", "Oracle products mapped to the architecture: an agent memory Python package that generates context cards; Deep Data Security, which propagates user identity and row-level access into the database so agents inherit the user's permissions; a Private AI Service Container for on-prem embeddings, LLM inference and re-rankers; OCI-hosted Grok and OpenAI models; and Private Agent Factory, a no-code drag-and-drop agent builder on Oracle AI Database." ], "takeaways": [ "Treat memory as data and own it, because 'memory is basically your moat'. Engineer the capture of memory traces, since they become fine-tuning data for continual learning.", "Write down your entities and definitions: build ontology and semantic layers so agents resolve domain terms correctly instead of relying on keyword search.", "Push entitlements into the data layer so an agent acts with the identity and row-level access of the user it serves.", "Look across your architecture for steps a small fine-tuned decision model could replace, such as memory routing, embeddings and re-ranking, to cut cost and latency.", "Measure agents on tokenomics: cost, accuracy and latency." ], "topics": [ "agent memory", "intelligence plane", "control plane", "continual learning", "ontology", "semantic layer", "fine-tuning", "data governance" ], "tools": [ "Oracle", "Oracle AI Database", "Oracle Deep Data Security", "Oracle Private AI Service Container", "Oracle Cloud", "Oracle Private Agent Factory", "Oracle agent memory Python package", "Memories (open-source, by Richmond Alake)", "Qwen", "Grok", "OpenAI", "General Electric" ], "quotes": [ { "text": "everyone in this room has the same models, probably the same access to the same information, but that doesn't equate to intelligence", "at": "07:21", "url": "https://www.youtube.com/watch?v=1_lX0m64-aI&t=441s" }, { "text": "information does not equal intelligence.", "at": "09:37", "url": "https://www.youtube.com/watch?v=1_lX0m64-aI&t=577s" }, { "text": "So memory is is basically your moat. You need to own your own memory.", "at": "17:27", "url": "https://www.youtube.com/watch?v=1_lX0m64-aI&t=1047s" }, { "text": "In 32,000 token compared to the 200 thousand token with the naive agent.", "at": "14:59", "url": "https://www.youtube.com/watch?v=1_lX0m64-aI&t=899s" } ], "words": 5255 }, "summary_url": "/#1_lX0m64-aI", "transcript": { "html": "/transcripts/1_lX0m64-aI.html", "txt": "/transcripts/1_lX0m64-aI.txt", "vtt": "/transcripts/1_lX0m64-aI.vtt" } }