模型都一样，智能体靠什么拉开差距：Oracle 的 Richmond Alake 谈记忆与智能平面

简要描述Brief Description

每个人都拥有相同的 AI 模型。智能体凭什么脱颖而出——Richmond Alake，OracleEveryone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake， Oracle

目录Table of Contents

• 简介
  Intro
• 一分钟了解 Agent 记忆
  Agent memory in one minute
• 控制平面
  The control plane
• 软件工厂
  The software factory
• 智能平面
  The intelligence plane
• AF 是中途岛
  AF is Midway
• 它在智能体技术栈中的位置
  Where it fits in the agent stack
• 情报层是什么
  What the intelligence plane is
• 智能体 = 模型 + 执行框架
  Agent = model + harness
• 为何是现在：决策模型
  Why now: decision models
• 演示：受治理的智能体与朴素智能体
  Demo: governed vs. naive agent
• 你们的团队现在应该做什么
  What your team should do now
• 智能体记忆、安全与私有推理
  Agent memory, security and private inference
• 演示：微调内存路由
  Demo: fine-tuning memory routing
• 总结：掌控你的记忆
  Wrap-up: own your memory

简介Intro

主持人：我接下来会讲一串关键词，大家也要积极互动。每讲到一个关键词，我都会请大家举手。听说过智能体记忆的，请举手。Host: I'm going to go through a bunch of keywords and you guys get interactive. I'm going to ask you guys to put your hands up as I go through a bunch of keywords. So, hands up if you heard about agent memory.

一分钟了解 Agent 记忆Agent memory in one minute

这让我开心得不得了，因为我之前做过这件事——大约一年前，我做过一场关于智能体记忆的演讲，当时在场的人没有一个知道我在说什么。那场演讲也是在旧金山的 AI Engineers 上做的。曾几何时，大概三年前，我谈到我们需要构建记忆结构、为智能体中的记忆建立模型时，大家都觉得我疯了。而现在，在座的各位都了解智能体记忆了，希望你们也正在智能体中实现它。所以，我就快速讲一下智能体记忆，真的很快。智能体记忆基本上就是智能体用来保留、重用、回忆和完善信息的机制。我把这称为“四个 R”，很容易记住。不过，智能体记忆也有一套分类体系。短期记忆和长期记忆比较容易理解。That makes me unbelievably happy because I did this I did a talk on agent memory about a year ago and no one in the room knew what I was talking about. So, it was also at AI Engineers San Francisco and once upon a time, maybe around 3 years ago, people thought I was crazy when I was talking about we need to structure memory, we need to model memory in agents. And now everyone here knows about agent memory and hopefully you're implementing it within your agent. So, I'm just going to quickly go over agent memory, just very quickly. So, agent memory is basically how an agent or the memory mechanism an agent uses to retain, reuse, recall, and refine information. I call it the four Rs, it's easy to remember. Um but there's also taxonomy of agent memory. Short-term, long-term is easy.

然后我们会讲到智能体记忆的认知功能：语义记忆、情景记忆、程序记忆和工作记忆。接着，我们会介绍一种今天会经常看到的实现方式。我已经在线上讲过很多次了，不过我们还是拍几张照片吧，而且大家也会拿到幻灯片。这就是记忆的分类体系，也是我一直想向大家普及、并且过去三年来一直在讲的内容。不过，在这场演讲中，我们还会更进一步。接下来又要出现一个关键词了。Then you go into the cognitive function of agent memory, semantic, episodic, procedural, working memory. Then we go into an implementation that we're going to see a lot today. And I've done this so many times online, but let's take pictures and you guys are going to get the slides as well. So, this is all the taxonomy of memory, the what I'm trying to educate people on and I've been educating people on for the past 3 years. But in this talk, we're going to take it further. So, another keyword coming up.

控制平面The control plane

听说过控制平面的举手？不错，举手的人少了些。举手的人越来越少了，我有个猜想。随着我们逐个讲解这些关键词，你会看到演示过程中举手的人数逐渐减少。要是一切顺利，最后就不会有人举手了，或者只剩一两个人。那么，控制平面可以这样定义。大多数人都知道它是什么。它本质上是一个集中式管理系统，用于管理你的智能体：它们受什么约束、可以使用哪些工具、能做什么以及不能做什么。很多软件公司都把自己的产品定位为控制平面。没错，我们确实需要这类系统。但在这场演讲中，我会带你再深入一点。这里先简单介绍一下控制平面是什么：智能体的身份、策略和护栏、编排与权限管理，以及所有审计功能。Hands up who's heard of the control plane? Nice, less so. We're getting We're getting less hands and I have a theory. As we go through the keywords, you're going to see the number of hands in the presentation reduce. Eventually, if things go correctly, there will be no hands or maybe one or so. So, the control plane, we can define it. Most people know what it is. It's essentially a centralized management system for how your agent is governed, with the tools they have access to, what they can do, and what they cannot do. So, there's a lot of different software companies positioning themselves as a control plane. And yes, we do need these sort of systems. But in this talk, I'm going to take you a little bit further. So, this is just an idea of what the control plane is. So, identity for your agents, the policy and guardrails, the orchestration to permission, and all the auditing.

这些就是典型控制平面的组成部分。好，接下来又要讲一个关键词了。大家准备好了吗？准备好了吗？好，开始。软件工厂。And these are the components of typical control planes. All right, another keyword is coming. You guys ready? You guys ready? All right, let's go. Software factory.

软件工厂The software factory

好，少举点手。我们快讲到了。现在讲的内容和我幻灯片上的内容几乎一致了。软件工厂其实是个挺有意思的术语。很久很久以前，我曾在通用电气实习过。当时在通用电气，有个人叫鲍勃·贝默（Bob Bemer），他在 1968 年的一篇论文中创造了“软件工厂”这个术语，探讨的是如何以可度量、可重复的方式创建程序。基本上，“软件工厂”这个术语就是他提出的。所以，它并不是智能体时代才出现的新术语，而是我们很久很久以前就在计算机科学领域使用的一个术语。我等大家拍完照。你可以把软件工厂想象成这样：你想创建一个能够可靠地持续产出软件的系统，对吧？你给它一套流程和系统化的工作流，让它在构建软件时遵循，然后让智能体来执行这些流程，通常是多个智能体并行工作。Okay, less hands. We're getting there. It's It's We're almost in line with what I have on the slides. So, software factory is actually an interesting term. So, I did my internship at General Electric like eons ago. And software factory is a term that someone called Bob Bemer at GE at the time in 1968, he coined the term software factory in one of his paper where he was talking about how we can create um programs in measurable and repeatable way. And essentially, he coined the term software factory. So, it's not a new term within the agentic era. It's a term that we've used in computer science a very, very long time ago. I'll wait for people to take pictures. So, software factory, you can just imagine it as you want to create a system that is able to churn out software reliably, right? You give it a bunch of um processes and systematic workflow it's meant to follow whenever it's building a software, and you have the agents execute this, usually multiple agents in parallel.

同样，不同的公司都在这个领域定位自己。好，再来看另一个关键词。Again, different companies position themselves in this space. All right, another keyword.

智能平面The intelligence plane

智能平面。有人举手吗？没有，我没成功，只有三个人举手。嗯，是的，今天我们要聊的就是智能平面。我要做一个承诺——去年我做过同样的承诺，当时我聊的是智能体记忆，大家还觉得我疯了。今年我要做的承诺还是一样：我今天要告诉你们的信息，我要说的所有内容，都会让你们比现在大多数人领先六个月。去年聊智能体记忆时，我做过这个承诺。今年聊智能平面，我也做出这个承诺。所以，等我明年回来聊智能平面时，大家都会举起手来。别、别拿我的话当真。那这个怎么样？AF 是中途岛。Intelligence plane. Any hands up? No, I failed. I got three hands up. Um yeah, the intelligence plane is what we're going to be talking about today. And I'm going to make a promise that I did in my previous talk last year when I was talking about agent memory when people thought I was crazy. And this promise is the same I'm going to I'm going to make, which is the information I'm going to give you today, everything I'm going to say, is going to put you 6 months ahead of where most people are today. So, I made that promise last year talking about agent memory. I'm making that promise this year talking about the intelligence plane. So, when I come back next year and I talk about the intelligence plane, everyone's going to put their hands up. Don't don't quote me. How about this one? AF is Midway.

AF 是中途岛。AF is Midway

有人听说过这件事吗？没有？好吧，那我们来简单讲一点历史。第二次世界大战期间，珍珠港遭到日本帝国海军的袭击，对吧？当时有好几个小组负责截获日本军方的通信。一共有三个小组，我们这里只讲其中两个。一个小组设在华盛顿，另一个设在珍珠港。还有一个代号 AF。AF 代表某个地方，而这对太平洋战争的最终结果至关重要。长话短说，这两个小组掌握了大量零碎信息，这些信息来自他们原本负责破译的日本海军部队的通信。他们收集到许多零散的信息。基本上，那个设在珍珠港的小组——一群人在地下室里，把大量信息和截获的纸质电报收集起来，保存了好几个月——成功破译了这些信息，并发现 AF 实际上是中途岛的代号。Who's heard about this? No? Okay, let's uh do a bit of history lesson. So, in World War II um in World War II uh Pearl Harbor was getting um attacked by the Japanese Imperial Navy, right? And there was a bunch of groups that were meant to be intercepting communication from the Japanese army. And there were three groups. So, we're only going to talk about two groups here. One group was based in Washington and one group was based in Pearl Harbor. And there was a code AF. Now, AF stood for something and it was very critical to the outcome of the actual um Pacific War itself. But to cut the long story short, these two groups had a bunch of fragmented information from communication from the Japanese Navy force that they were they were that they were meant to codebreak. So, they had a lot of fragmented information and basically the one group, which is the group based in Pearl Harbor, a bunch of folks that were basically in a basement collecting loads of information and intercepted messages on paper and storing it for over months, were able to decode this and they found out that AF is actually short for Midway.

对吧？AF 就是中途岛。对吧？所以，日本人用 AF 作为中途岛的代号。中途岛是一座位于亚洲和北美之间的小岛。所以，这是……你们可以查一下，关系到太平洋的控制权。不过，我想告诉你们一个关键点，那就是控制权和情报之间的区别。因为珍珠港有一小群人，他们待在地下室里，破译所有的日本通信，还连续几个月用大量纸张记录一切——这本质上就是记忆——而他们实际上能够解读这些信息，让美国海军处于有利位置。相比之下，华盛顿的组织更集中，掌握的信息却是一样的，对吧？信息并没有区别。他们掌握着相同的信息，资源可能还更多，而且他们拥有控制权，也有组织架构，却对这些数据得出了完全不同的解读。Right? AF is Midway. Right? So, the Japanese were using AF as code word for Midway. Midway was a little island located in between Asia and North America. So, this was You can look this up, control of the Pacific. But, a key thing I wanted to tell you is the difference between control and intelligence. Because that small group that was in Pearl Harbor, located in a basement, and they were deciphering all the um Japanese communication, and using a lot of papers to record everything for months, which is essentially memory, what was able to actually decode and position um the American Navy at an advantage. So, whereas Washington, which was more centralized, had the same information, right? There was no difference in information. They had the same information, probably more resources, and they had control, and they had organization, came up with an entirely different um uh interpretation of the data.

最终被选中的，是我们夏威夷站 Hypo 的那群人，也就是那里的密码破译人员。他们之所以也能够提出我们的假设——AF 确实代表中途岛——就是这个原因。我想说的是，在座的每个人都有相同的模型，也可能能访问相同的信息，但这并不等同于智能，也不意味着你拥有那种独特的差异化优势。所以，今天我要谈谈，为什么我们真的需要开始讨论智能层。And the one that was chosen was the folks of our Station Hypo, um the code breakers there. And the reason they were able to also do our hypothesis on the fact that AF does mean Mid- Midway. So, what I'm trying to say is everyone in this room has the same models, probably the same access to the same information, but that doesn't equate to intelligence, and it doesn't equate to you having that unique differentiator. So, today, I'm going to talk about why we really need to start talking about the intelligence plane.

它在智能体技术栈中的位置Where it fits in the agent stack

那么，我要来梳理一下智能体栈。我之前谈过智能体栈，我们知道它是什么，而控制平面会贯穿智能体栈的某个层级。现在，我想向大家提出的并不是什么新东西，只是提供一种思考方式，帮助大家看待当下正在构建的软件。这就是智能平面。它贯穿整个技术栈，开辟出一个维度，让你们作为工程师能够开始构建智能体——它们不仅可以被编排，还能够从数据和记忆中学习。那么，我们接下来要从理论转向更实际的概念。So, I'm going to go through the agent stack. I've spoken about the agent stack. We know what the agent stack is, and then the control plane cuts through a certain level of the agent stack. Now, what I'm proposing to you is nothing new. It's just a way of thinking about the software you're building today. So, this is the intelligence plane. It cuts through the entire stack, and creates a dimension where you as engineers can start to build agents that are not only orchestrated, but they can actually learn from data and memory as well. So, we're going to we're going to move from theory into more practical concepts.

什么是智能平面What the intelligence plane is

那么，什么是智能平面？它本质上是智能体系统或 AI 基础设施的一部分，能让你了解情况，或让智能体随着时间推移不断进步。它涵盖了智能体应该知道什么——也就是知识；它应该记住什么——也就是记忆；它可以看到什么——也就是治理和数据访问；以及它如何不断进步——也就是持续学习。所以，这里并没有什么新东西。你们当中有些人已经在做其中的一些事情，有些人听完这场演讲后就会开始尝试。这就是控制平面和智能平面的区别。大家跟上了吗？对吧？点点头？很好，我们继续。嗯，我会带大家回顾这条时间线，看看大多数人已经走到哪里，以及接下来会走向哪里。So, what is the intelligence plane? It's essentially a part of your agentic system or your AI infrastructure that allows you to understand or allows your agent to get better over time. It encapsulate what the agent should know, which is basically knowledge, what it should remember, which is basically memory, what it's allowed to see, which is just governance, data access, and how it gets better, which is continual learning. So, there's nothing new here. Some of you are already doing some of these things already, and some of you are going to start doing it after this talk. So, that is the difference between the control plane and the intelligence plane. Are we following? Right? Nodding? Okay, this is good. Let's go. Um I'm going to take you through this timeline of where most of you have been and where most of you are going to get to.

首先，我们都在给模型输入提示词，而且我们都有类似的提示词层或工具，可以查看所有提示词。然后，我们转向了能够让我们查看上下文，也就是上下文层的工具。现在，你们大多数人都处于控制平面时代：你们正在寻找或构建能够编排这些智能体的软件。接下来，我们要迈向的是智能平面；或者说，你们应该朝智能平面发展。 先说明一点，在所有这些幻灯片中，有一件事我们可以理解并达成共识：信息不等于智能。记住这一点，现在我们可以开始深入一些技术内容了。 这就是智能体技术栈。我们谈过智能体平面。现在我们来谈谈控制平面。当我们开始思考如何在控制平面中融入智能，或者为控制平面增加智能时，如果从网关和连接能力来看，就会开始关注与数据相连的端到端用户身份。First, we were all prompting the models, and we all had like prompt layers or tools where we can see all our prompts. Then we moved into tools that can allow us to see the context, context layer. And most of you are now in the control plane error, when you guys are looking for softwares or building softwares that allows you to orchestrate this agent. So, where we're going to go to is intelligence plane or where you should be heading to is intelligence plane. And just to get this out there, throughout all the slides, one thing we can understand and agree on, information does not equal intelligence. So, take that away, and now we can start getting a bit technical. So, this is the agent stack. We spoke about the agent plane. And when we start looking about let's talk about the control plane. When we start thinking about intelligence in the control plane or intelligence added onto the control plane, we start to think about when you're looking at gateway and our connectivity, you start to think about end-to-end user identity connected to the data.

对吧？当你开始思考编排时，就会开始考虑上下文组装，也会想到记忆和上下文构建等事情。接着一直讲到记忆核心。我直接跳到记忆核心，因为这是我的专长——数据库。嗯，我们开始讨论数据库，但还会进一步探讨本体、语义层以及记忆。然后，基础设施和平台才是重点。接下来我们会开始讨论推理、微调和持续学习，对吧？这就超出了控制平面提供的部署能力。因此，这些都是控制平面和智能平面的组成部分。我们会提供一套参考架构，你可以把它拍下来。Right? When you start to think about orchestration, you start to think about the context assembly, you start to think about things such as memory and contest construction. Then all the way to the memory core. I'm skipping to the memory core because that's my sweet spot, database. Um we start to talk about database, but we take things further. We start to talk about ontology and semantic layers and memory as well. Then infrastructure and platform is the important bit. Which is where we start to talk about inference, fine-tuning, and continuous learning, right? When we go beyond the deployment that the control plane gives you. So, these are all the components of the control plane and the intelligence plane. There's going to be a reference architecture that you would be able to take a picture of.

这个想法嘛，我希望是我想出来的新点子，但我现在向你们提出的这个想法其实并不新。嗯，它已经被提出过好几次了。事实上，如果你们有人做过网络相关的工作，可能会觉得这个想法很熟悉。嗯，有些人提出过，在管理和配置平台之上，应该有一个知识平面，用来实际查看不同电路的路由——这本质上就是一个控制平面，对吧？所以，有控制平面、管理平面，后来又有人提出了知识平面。而这就是我为我们正在构建的智能体系统提出的方案：一个智能平面。我们会做一个演示。我看看时间，确保能讲到演示。我们也喜欢方程式。And this idea, I wish I came up with new ideas, but this idea I'm actually proposing to you is nothing new. Um it's been proposed several times. In fact, in if you If anyone's done networking, this might be familiar to you. Um some folks have proposed that there needs to be a knowledge plane on top of management and configuration platforms that are meant to actually see um the the routes of different circuits, which is essentially a a control plane, right? So, a control plane, a management plane, then folks have um uh proposed a knowledge plane. And this is what I'm proposing for the agentic systems we're building, an intelligence plane. We're going to have a demo. I'm looking at the time to make sure I get to the demo. And we like equations.

智能体 = 模型 + 执行框架Agent = model + harness

智能体等于模型加 harness。然后把它纳入控制，也就是治理加安全。但当你引入智能能力，加入知识、记忆和学习，就能在你如今构建智能体的整个过程中产生乘数效应。我来给大家展示一下具体效果。我在想，现在要不要做演示，还是晚点再做？嗯，我们先做个演示版。哦，我先回到前面。刚才大家都在拍照。马上就会把幻灯片发给大家。好。那么，我提到的智能平面中的所有内容，都尽量概括成了高层级的系统组件。现在，这些组件分别可以由不同的提供商提供，对吧？首先是智能体运行时，你的智能体将在其中执行。接着，智能体运行时需要记忆和知识；它可能还会有一个构建在数据平台之上的记忆管理器。然后，智能体运行时可以将记忆痕迹写入记忆管理器，之后这些内容还可以进入持续学习循环，从而让你能够更新模型，并将其部署回智能体运行时。Agent equals model plus harness. Then you take that into control, that equals governance plus security. But when you bring intelligence, you add knowledge and memory and learning, it gives you that multiplier effect on your entire on the entire way you're building agents today. And I'm going to show you what that looks like. I'm trying to think, should I go to the demo now or later? Um we'll do a demo version. Oh, I'll go back. People were taking pictures. You are going to get the slides um in a second. All right. So, everything I talk about within the intelligence plane, I've tried to capture as just high-level system components. So, all of these would be different providers today. Right? So, we have the agent runtime that your agent is going to be executing on. Then your agent runtime needs memory and knowledge, probably has a memory manager on top that sits on top of data platform, and then your agent runtime is able to put in traces of memory into your memory manager, and that could then go into a continual learning loop that allows you to then change models back into your agent runtime.

对吧？你可以把这一切都放到我们一直在讨论的这些平面上。所以，我们有数据平面、控制平面和智能平面。我就不展开讲了，因为我们已经知道控制平面是什么，也知道智能平面是什么，而智能平面正是缺失的关键部分。在智能平面中，有几件事非常重要。我讲得有点太快了。接下来我会通过一个演示展示其中几件非常重要的事。这就是我马上要展示的内容。那么，为什么这个行业已经准备好迎接智能平面了？Right? And you can put this all on the plane that we've been talking about. So, we have the data plane, the control plane, the intelligence plane. I'm not going to go over this cuz we already know what the control plane is. We know what the intelligence plane is, and that's the key thing that is missing. So, in the intelligence plane, we have a couple things that are actually quite important. I'm going a bit too fast. We have a couple things that are quite important I'm going to show in a demo. And that is what I'm going to be showing in a second. So, why is the industry ready for the intelligence plane?

为什么是现在：决策模型Why now: decision models

为什么我现在要谈智能平面？简单来说，我们有前沿 AI 模型。嗯，规模定律让我们能够预测从这些模型中可以获得哪些收益，而开放权重模型通常会落后大约 6 个月或 3 个月。不过，最近发生了一件事，那就是决策模型，对吧？它们改变了发展轨迹，也改变了我们开始思考这些系统的方式。我对这些系统的思考方式，对吧？这让我开始从非常基础的架构层面思考这些系统。审视我的系统，并思考：好吧，这里有哪些部分可以用决策模型来替代？几年前我们看待事物的传统方式是这样的，而如今，你可以在整个检索流程中使用模型。所以，我接下来要做的一件事，就是在这里开始演示。Why am I talking about the intelligence plane now? So, very quickly, we have frontier AI models. Um the scaling law allowed us to actually predict the benefits we could get from these models, and then we have the open ways are very like 6 months or 3 months behind. But we have we had something that happened recently, which was decision models. Right? And they changed the trajectory or the way that we started to think of these systems. The way I thought about the systems, right? It made me start to think systems that are very fundamental architectural level. Looking at my system and thinking, okay, what here could I replace as a decision model? So, this was the traditional way we were seeing things a few years ago, and today you can have models all through your retrieval pipeline. So, one thing I'm going to do is I'm going to go into a demo right here.

这是演示用的二维码。This is the QR for the demo.

演示：受治理的智能体与朴素智能体Demo: governed vs. naive agent

我有点说得太超前了。我们开始吧。在这个演示中，你们会拿到代码，不过我不会现场运行，因为这里的网络不太好，你们应该也知道。嗯，这是一个简单的聊天界面。我有一个治理型分析代理，还有一个朴素型分析代理。我的治理型分析代理构建在智能平面之上。智能平面为它提供本体层、语义层，以及我们在参考架构中看到的所有功能。简单演示一下智能平面带来的效果。关键在于拥有这样一个智能平面，从而节省 token 开销，对吧？Token 开销关系到成本、准确性和延迟。我们有两个代理能够执行相同的问题，但由智能平面驱动的那个更聪明一些，99% 的时候都能答对。A bit going ahead of myself. Here we go. So, in this demo, which you're going to get the code to, I'm not going to run it live cuz the internet here isn't great as you guys probably know. Um this is a simple chat interface. I have a governed analyst, I have a naive analyst. My governed analyst is built on what is an intelligence plane. It gives it the ontology layer, semantic layer, gives it everything we've seen in that reference architecture. And very quickly, just to show you the effect of having an intelligence plane. The key thing is having one, saving on tokenomics, right? Tokenomics is cost, accuracy, and latency. And we have two agents that are able to execute the same questions, but the one that is powered by an intelligence plane is a bit smarter. It gets things right 99% of the time.

而且它消耗的 token 更少：只用了 32,000 个 token，而朴素代理用了 20 万个。我们还设置了所有防护控制，也配备了缓存机制。此外，我们还有一些提示词缓存机制。所以，等你真正开始——我这里是倒着讲的——等你开始实现我马上要给你看的这些东西时，你就会开始构建这样的系统：第一，能够针对生产环境进行优化；第二，能够利用你所采集的数据不断学习和成长。这里我想谈谈本体和语义层。我在几个演示之间来回切换，因为时间不多了。关于本体和语义层，我向两个代理都问了一个问题：什么是交易对手方？And it does it with less token consumption. In 32,000 token compared to the 200 thousand token with the naive agent. And then we have all of the guard controls in place and the cache as well. And we have some prompt caching mechanism as well. So, when you actually start I'm working backwards here. When you start to implement some of the things I'm going to show you to you in a second, you start to implement systems that are able to actually one, be optimized for production, and two, learn and grow with the data that you're capturing. So, over here I just want to talk about the ontology and the semantic layer. I'm kind of skipping between demos here cuz we're short on time. So, with the ontology and semantic layer, I asked the question uh to both agents around um what was a counterparty, right?

本体层带来了更高的准确性，这是单靠简单的关键词搜索无法实现的。在演示中，这是一次关键词搜索，我们确实能够找到一些实体，对吧？没有本体时，智能体也能回答，但加入本体后，答案确实有所改善，并且借助本体得出了正确答案。语义层也是如此。我们向一个没有任何语义层的朴素智能体提了个问题：浓度是什么意思？它答错了。而由语义层和智能层构建的智能体则答对了。现在，误差率只有 5%。但在金融领域，任何关键任务工作负载都容不得差错，这就变得至关重要。And the ontology layer adds a level of accuracy that you can't get by just doing simple keyword search. So, in the demo, this is a keyword search where we're able to actually find some um entities, right? And the agent was able to answer without the ontology, but with the ontology, the answer actually improves and it gets the right answer using the ontology. Same with the semantic layer. The agent the naive agent without any semantic layer, we asked it a question and it got the we asked it what does concentration mean? And it got it wrong. The naive agent got it wrong, whereas the semantic the intelligent agent built with the semantic play with the intelligence plane got it right. And now that margin of error is just five 5%. But when you're working with the finance sector in any mission critical workload, that becomes very critical.

因此，构建好你的语义层和本体层变得非常重要。你们将会获得这个演示的访问权限，我现在要很快回到幻灯片上。我们正在将一系列模型添加到智能平面中，而我刚才给你们展示的演示实际上已经部署了所有这些模型和整套架构。So getting your semantic layer built out, get your ontology layer built out, it becomes very important. I'm going to you're going to get access to this demo and I'm going to go back to the slides very quickly. So there's a bunch of models that we're adding into the intelligence plane and the demo I showed you actually has all of these models and all of this architecture in place.

你的团队现在应该做什么What your team should do now

那么，你们团队现在应该做什么？第一，把记忆视为数据，对吧？不要把它看成彼此不同的东西。把你们的实体和定义记录下来。把权限推送到数据层。在 Oracle 上，我们提供了一些功能，让你们能够非常轻松地做到这一点并捕获追踪记录。所以，记忆基本上就是你们的护城河。你们需要拥有自己的记忆。正如我们之前看到的，在 Oracle 上，我们有这样的参考架构。而在 Oracle，我们也有与之对应的产品功能。因此，我们为你们提供构建自己的智能平面所需的系统组件。这意味着，你们拥有所需的所有工具，可以构建能够不断学习和成长的系统；随着你们持续构建并运行智能体，它也会不断学习和成长。接下来我会谈谈其中几个方面：记忆、安全、深度数据安全、私有推理，以及本体和语义。So what should you your team be doing now? One, you treat memory as data. Right? Don't look at it as different things. Write down your entities and definition. You push entitlements into the data layer. Over Oracle, we have features that actually enable you to do this very easily and capture traces. So memory is is basically your moat. You need to own your own memory. So over Oracle, we have this reference architecture we saw before. But at Oracle, we have product features that map to it. So we give you the system components to build your own intelligence plane. So that means that you have all the tools you need to build something that actually learns and grow as you continue to execute build as you continue to build and execute your agent. I'm going to touch on a few ones. Memory, security, deep data security, private inference and ontology and semantics.

因此，以下是我想快速谈及的五个关键领域。So these are the key five areas I want to touch on very quickly.

智能体记忆、安全与私有推理Agent memory, security and private inference

说到智能体记忆。我们之前已经聊过这个话题，而 Oracle 现在做的，是将智能体记忆作为一个 Python 包推出。我们把记忆工程封装起来，这样你就不必操心让智能体记住信息所涉及的各种复杂问题；我们将它做成一个 Python 包，由它为你生成上下文卡片，你无需亲自处理。接下来，你需要开始采集记忆轨迹，因为这些数据会用于微调。这里有一张幻灯片，展示了记忆工程与上下文工程的交汇之处。我把幻灯片发给你后，你可以仔细看看。我先快速略过这一部分。我们刚才在演示中介绍的，就是智能体记忆和 Oracle AI 数据库。So agent memory. We've spoken about it, but now what we did at Oracle is we launched agent memory as a Python package. We put memory engineering into a box, which allows you to just not worry about all of the complexities of making agents remember, we just convert it into a Python package and it generates context card for you where you don't have to actually do it yourself. Then, one thing you have to actually have to start capturing is your memory traces because this become data for fine-tuning. So, right here we have a slide that just shows where memory engineering and context engineering meets and this is for you to take and go through when I give you the slides. I'm just going to skip through this very quickly. So, agent memory and the Oracle AI database is what we just covered in the demo.

我跳过了。不过，接下来要讲的是深度数据安全。昨天和今天我跟几位朋友聊过，他们谈到了 Agent 身份，以及他们如何使用不同的解决方案提供商来处理这个问题。在 Oracle 的 Agent 中，我们做的一件事，就是将安全作为数据库的一等公民。也就是说，通过一款名为“深度数据安全”的产品，你可以获取应用程序的所有上下文，以及用户的数据访问权限和行级访问权限，并将这些信息传递到数据层。这意味着，你的 Agent 可以采用它所代表的用户的身份，并访问与该用户相同的数据——也就是说，只能访问该用户获准访问的数据。我刚才有点语无伦次。好。那么，所有这些都很棒，包括我给你们看的演示，它用到了我刚才提到并展示的所有这些基础设施。I skipped it. But, the next thing is deep data security. I spoke to a few folks yesterday and today and they were talking about agent identity and how they were using different solution providers for it. In agent in Oracle, one thing we did is we made security a first-class citizen of the database. Which means with a product called deep data security, you can take all of the context of your application and the data access and row access of your user and we propagate that into the data layer. Which means your agents can take on the identity of the user it's acting on behalf and access the same data that it's only allowed for is only that the user is only allowed to access. Fumbling my words there. Okay. So, all of this is all great including the demo I showed you it uses all of this infrastructure I spoke about and I showed you.

这就是深层数据安全的工作方式。没错，我们有分析师，也有身份提供商。身份提供商会提供分析师应有的所有数据访问权限，我们可以将这些权限传递给智能体，让智能体也继承这些权限。这样就能限制该智能体对所有数据的访问。好，接下来的几张幻灯片，我要讲讲私有推理。对吧？你们很多人都在金融公司工作，希望能在本地部署推理服务。因此，我们提供私有 AI 服务容器，其中包含托管自有嵌入模型、自有 LLM、自有 LLM 推理服务以及重排序器所需的全部基础设施。这一切都是为了让你拥有自己的智能平面。我们会为你提供这些组件。不过，如果你想转向云端，我们也提供 Oracle Cloud，并在其中托管一些你们正在听说的模型。And this is how deep data security works. That's right. We have the analyst, we have our identity provider. The identity provider provides all of the data access that the analyst is meant to have and we can propagate that into the agent and the agent inherits that as well. And that limits all of the data access for that particular agent. Okay. The next couple of slides I'm going to talk about private inference. Right? A lot of you work in finance companies where you want to actually have your inference on prem. So, we have the private AI service container which contains all the infrastructure you need to host your own your own embedding, your own LLM in your own LLM inference, and re-rankers. And this is all around owning your own intelligence plane. So, giving you the components. But, if you want to go to the cloud, we have Oracle Cloud where we host some of these models that you're hearing about.

所以，Grok 模型以及一些 OpenAI 模型也一样。无论是物理隔离的基础设施还是云基础设施，Oracle 本身都具备这些组件。So, Grok models and some Open AI models as well. Either way, air-gapped um infrastructure or cloud infrastructure, you have the components within Oracle itself.

演示：微调内存路由Demo: fine-tuning memory routing

好。在结束之前，我还有第二个演示。这个演示介绍的是我开发的一款名为 Memories 的开源工具，主要是想向大家展示如何开始进入智能平面领域。目前，已经有人在研究智能体记忆，正如大家提到的那样。而在智能体记忆系统中，还有其他方式可以引入模型来改进系统本身。决策模型就是其中一种，这个演示会用到三种不同的模型：决策模型、嵌入模型和重排序模型。在我的智能体记忆系统中，我通常会使用这些模型。不过，借助智能平面，或者说，认识到我需要拥有自己的记忆，并且确实由自己掌控记忆，我就可以开始对其中一些开源模型进行微调。All right. So, I have a second demo before I go off. So, this demo is an open-source tool that I built called Memories. And it's just to show you what how you can actually start to get into the intelligence plane domain. So, right now there are people working on agent memory, like you guys have mentioned. And within an agent memory system, there are other ways you can start to bring in models to improve the system itself. So, decision model is one of it, and this one makes exuses example of three different models. A decision model, an embedding model, and a re-ranker. And within my agent memory system, I typically have these models. But, with the intelligence plane or with an understanding that I need to own my own memory, and I own my own memory, I can actually start to fine-tune some of these open-source models.

这就是我们在这里要做的事情。现在我们来看一下这里的记忆路由，我先简单讲讲记忆路由。所谓记忆路由，就是我们从上下文层或 LLM 获取一些输出，然后将其路由到正确的存储区。无论输出是什么，我都可以把它路由到对话记忆或工作流记忆中。通常，你需要通过编写代码来实现这一点，但借助决策模型，这件事会变得非常简单，而且成本很低。现在，我们这里有一些经过微调、用于记忆路由的模型，可以看看它是如何训练的。我们对 Equine 模型进行了微调，并将其训练成一个决策模型。用于微调的数据都在这里，取自我与智能体的对话；我能够据此判断记忆的类型。And that's what we're doing here. So, over here in memory routing, I'm just going to focus on memory routing for a second. And what memory routing means is we get some output from the context layer, or we get some output from our LLM, and what we're able to do is route it to the right store. So, whatever comes out, I can route it into conversation memory or workflow memory. Normally, you would do this using some code, but with decision models, this becomes very easy and very cheap. So, right here we have some fine-tuned Oh, skipped it. Let's go back. All right. Right here we have some fine-tuned models for memory routing, and we can see how it was trained. So we fine tune the equine model and we use it and we fine tune it into a decision model. The data that we use to fine tune is all here and obtained from the conversation I was have been with the agent and I was able to classify what type of memory it was.

这样就能非常轻松地实现记忆路由。一旦有信息进入，你就可以使用决策模型，根据不同类型的记忆，以高效、低成本的方式将信息路由到相应位置。记忆路由、嵌入和重排序，是你们今天就可以着手尝试的一些最容易实现的改进点。了解了智能平面之后，你们就可以捕获记忆轨迹、有效地对其进行格式化，并重新训练模型。这样既能提升质量，有时也能降低延迟。好，我们再回到幻灯片。我们来回切换了几次，这样挺好。嗯，我们就此做个总结。在 Oracle，我们正通过多种方式帮助开发者构建智能体。And this allows you to do memory routing very easily. So once you get any information coming in, the different types of memory you want to route it effectively and cheaply, you can do that with decision models. So memory routing, embeddings, and re-ranking are some of the lowest hanging fruits you guys can actually start to work with today. Once you understand your intelligence plane, you're capturing your memory traces, you're formatting them effectively, and you're retraining the models. And you get the benefit with improvement in quality and in some cases reduction in latency as well. Okay, so we're going to go back to the slide again. We're going back and forth a few times, which is good. Um and we're just going to wrap up. So there are different ways we're helping developers build agents in in Oracle.

其中之一是私有智能体工厂。这是一个无代码平台，你只需拖放组件就能构建智能体，并在私有智能体工厂内全面掌控整个控制平面。私有智能体工厂也采用了我们一直在各个产品界面中构建的同一套智能体记忆系统。再说一遍，私有智能体工厂构建在 Oracle AI 数据库之上，因此你可以获得我刚才提到的所有记忆核心功能。我已经展示过本体和语义层，以及它们对智能体准确性的影响——有本体和语义层与没有它们的智能体，准确性会有什么不同。不过，Oracle 会为你提供构建这一切所需的所有基础设施和组件。因此，你可以用有意义和没有意义的方式提出同一个问题，对吧？我们已经做过演示了。And one of it is private agent factory. And this is a no code platform we can just drag and drop and build your agents and have the whole control plane at least within the private agent factory. And the private agent factory also has the same agent memory system that we've we've been building across all our product surfaces. So again, the private agent factory is built on top of the Oracle AI database, so you get all of the memory core functionalities I've been talking about. I already showed the ontology and semantic layer the effect it has on the accuracy that you get within agents that have and don't have ontology and semantic layer. But again, Oracle gives you all the infrastructure and components to build all of this. So you can ask the same question with and without meaning, right? So we've already done the demo.

再说一次，我略过了……我基本上是在快速浏览幻灯片和演示，不过你们已经看到了我们在有本体层和没有本体层时的表现。准确率的提升可能微乎其微，但在任务关键型工作负载中，这可能意味着会不会因为处理欺诈交易而被罚款一亿美元。Again, I skipped I'm kind of like going around the slides and the demos, but you saw what we had with ontology layers and without ontology layers and the accuracy improvements can be very minimal, but in mission-critical workload, they can mean getting a 100 million fine for handling fraudulent transactions or not.

总结：掌控你的记忆Wrap-up: own your memory

好的，我们刚才简单谈到了持续学习。演示中展示的就是持续学习。对于持续学习，有不同的观察方式：可以从 token 空间、权重空间来看，也可以从潜在空间来看，尤其是在涉及提示缓存时。我们刚刚已经讲过这些内容，这里就不再赘述了。不过，关键的一点、也是最重要的启示是：作为这些系统的开发者和工程师，我们需要真正掌握自己的记忆机制，精心设计轨迹的采集方式，并将其构建到系统中，让持续学习成为可能。 我就讲到这里。我想我已经超时了。接下来给大家看一下刚才演示的二维码。大家可以扫码；不过，最好的方式是来 Oracle 展位找我，到时你们可以再次拿到这个二维码，以及其他技术资料，我们也可以聊一聊。Okay, so we touched on continual learning a bit. What we saw in the demo was continual learning. And which is there different ways you can look at continual learning in token space and in weight space and also in in latent space as well if you're going into the prompts caching. So we've just gone over this. I'm not going to rehash this again. But one of the key thing and the key takeaway here is is developers and engineers of these systems, we need to actually own our own memory, engineer the capturing of traces, and actually build it into a system that allows for continual learning to be possible. So I'm going to end right here. Um I think I'm going over time and I'm going to give you guys the QR code to the demo I just showed. So if you just scan this, but the best thing is you can come meet me at the Oracle booth and you will get this QR code again along with other technical assets and we can have a conversation.

感谢您抽出时间。Thank you for your time.