# 高性能存储加速生成式 AI 全生命周期 > **合规声明（生成式 AI 服务）**：前述特定亚马逊云科技生成式人工智能相关的服务目前在亚马逊云科技海外区域可用。亚马逊云科技中国区域相关云服务由西云数据和光环新网运营，具体信息以中国区域官网为准。 > > **完整资料：如果想了解更多内容可以访问:https://events.amazoncloud.cn/china-summit-on-demand/ > > **免责声明**：本摘要仅为便于您了解峰会内容而提供，我们不对其完整性及准确性作出任何保证，且不构成任何建议或结论。相关内容仅供参考，实际内容以峰会回放视频为准。 ## 一、基础信息 - **会议类型**：专题演讲 - **Persona**：数据团队 - **时间信息**：6月23日 | 15:00 - 15:30 - **标题**：高性能存储加速生成式 AI 全生命周期 - **PDF 资料**：有 - **视频回放**：有 ## 演讲人信息 ### 会议信息 - **地点**：上海世博中心（2026 亚马逊云科技中国峰会） - **日期**：2026年6月23日（Day 1） - **时间**：15:00 - 15:30 - **会议类型**：专题演讲 - **面向 Persona**：数据团队 ### 亚马逊云科技演讲人 - **Christian Smith**：亚马逊云科技全球存储业务拓展总监 ## 二、会议资料 PDF 转录内容 --- Page 1 --- 2026年6月23日-24日 上海·世博中心 --- Page 2 --- 高性能存储加速生成式 AI Accelerating Generative AI with High-Performance Storage Christian Smith 亚马逊云科技 WW Specialist Storage --- Page 3 --- 数据是生成式 AI 的关键差异化优势 Your data is the differentiator for Generative AI 存储策略对生成式 AI 至关重要 • 生成式 AI 从非结构化数据中释放新的业务价值 • 您的数据才是关键差异化优势 ——而非模型 • 无论是构建、定制还是训练模型，数据可访问性决定了结果 • 高质量数据的存储策略至关重要 • S3 是数据存储的基石：全球百万客户的数据湖构建在 S3 上 --- Page 4 --- 生成式 AI 生命周期：存储的关键作用 Generative AI Lifecycle: Where Storage Matters 01 02 03 训练 微调 / 优化 智能体 利用海量数据集训练基础模型 利用行业数据调整预训练模型以提 使智能体能够从领域特定数据中提 （万亿级 token,PB 级 升模型性能 取洞察，从而简化复杂任务 数据，数千个 GPU 实例） （LoRA、RLHF、全量微调） --- Page 5 --- 面向生成式 AI 的存储 Storage for Generative AI S3 FSx S3+ Amazon S3 / S3 Amazon FSx S3 Vectors / S3 Files / Express One Zone for Lustre Mountpoint S3 Vectors: 适用于大规模数据湖和模型训练的可扩 完全托管的并行文件系统 适用于 RAG 工作负载的向量存储 展对象存储 POSIX 语义兼容 S3 Files: S3 Express One Zone 提供 亚毫秒级延迟 通过 NFS/POSIX 访问 S3 数据个位数毫秒级延迟 高达 1 TB/s 的聚合吞吐 大规模人工智能场景（>1000 GPU） 适用于大规模分布式训练 Mountpoint: 为 ML 框架提供文件访问 无限容量 1 TB/s 吞吐量 向量搜索 10 倍速度提升（Express） 亚毫秒级延迟 NFS 访问 >1000 GPU 规模 POSIX 兼容 框架原生 --- Page 6 --- 分布式模型训练和调优的存储挑战 Distributed Training: Storage Challenges --- Page 7 --- 为什么存储性能对模型训练调优至关重要 Why Storage Matters 模型调优在您的领域特定数据上重新训练基础模型，存储性能决定了您 $32/小时 昂贵的 GPU 是保持忙碌还是闲置 GPU 闲置小时的成本 数据吞吐必须匹配 GPU 计算能力， 如果存储无法足够快地向 GPU 提 供数据，您就在为闲置的 GPU 资源买单 #1 存储性能瓶颈是模型训练中的主要挑战 错误的存储选择 = 闲置的 GPU = 浪费的支出 --- Page 8 --- Amazon S3 最佳适用场景 分布式模型训练/调优， 1-16 个实例的集群规模 能力 大规模数据场景 • 流式吞吐量 • 几乎无限的容量 • 按用量付费 • 与 SageMaker 和 PyTorch 原生集成 • 适用于：LoRA/QLoRA、模型调优 • 与 S3 Files 和 S3 Mountpoints 集成 挑战 小文件工作负载（亿级小文件） 频繁的检查点保存和恢复 --- Page 9 --- 最佳适用场景 FSx for Lustre 分布式模型训练/调优，16-2000 个实例的集群 能力 大规模训练场景 • 跨数千个客户端的并行 I/O • 亚毫秒级延迟，保持 GPU 满负荷运行 • 扩展至 1+ TB/s 聚合吞吐量 • POSIX 兼容(PyTorch、DeepSpeed、Megatron) • 与 S3 集成， 支持从 S3 延迟加载，无需手动管理数据复制和同步 • LZ4 压缩 + 原生 S3 检查点 • EFA+ NVIDIA GPU Direct Storage 支持 • 适用于：多节点大规模分布式训练 挑战 数十亿小文件（元数据服务器）场景，需要针对性的部署和规划 --- Page 10 --- S3 Express 最佳适用场景 超大规模场景—数万个实例 One Zone 适合于云原生且偏好对象存储接口的数据科学团队 能力 超大规模训练 • 低延迟， 比标准 S3 低 10 倍以上 • 大规模并发下的一致高吞吐 • 数据并行访问，线性扩展 • 为机器学习优化的存储桶 • 适用于：超大规模集群，1000+ 或更多的 GPU 集群 挑战 需要使用 S3 API 与标准 S3 存储桶没有直接的数据集成，需要手动管理数据迁移 --- Page 11 --- 通用建议：从 FSx for Lustre 开始 Customers typically start with FSx for Lustre • 简单易用：工具和数据加载器与文件系统原生集成 • 高性能：吞吐量随文件系统大小线性扩展 • 存储选择：持久化 vs. 临时： 按照工作负载生命周期来选择持久化存储或者临时存储 • 建议：从吞吐量需求开始规划，而非容量 • 成本优化：使用 Intelligent-Tiering 优化成本而不牺牲性能 --- Page 12 --- Adobe 仅用九个月就发布了 Firefly 生成式 AI 模型 利用 Amazon FSx for Lustre 加速大规模模型训练 — Adobe Creative Cloud & AI --- Page 13 --- 决策框架 Decision Framework 为训练选择合适的存储 就地使用您的数据 < 16 个实例 S3、EFS、EBS、FSxZ、FSxN 等 —— 无需移动数据 16-500:FSx for Lustre（除非团队或应用程序偏好对象存储） 16 - 1,000 个实例 500-1,000:FSx for Lustre 或 S3 Express OneZone 需要频繁检查点或 POSIX？→FSx for Lustre | 否则→团队对文件 vs 对象的偏好 S3 Express OneZone > 1,000个实例 在此规模下，经济性通常超越其他考量因素 --- Page 14 --- 面向 Agentic 工作负载的存储服务 Storage for Agentic Workloads --- Page 15 --- AI 通过 Agent 变得更加自动化 AI gets more autonomous with agents 经典机器学习 生成式 AI 助手 生成式 AI 智能体 Agentic AI 系统 更多的 更少的 人工监督 人工监督 分析任务 内容生成 更复杂的功能 自动化决策 预定义规则 有限的决策功能 和任务 高度自适应性 有限的适应性 MCP 支持 --- Page 16 --- Agents 需要多种类型的数据 短期 长期 S3 Express S3 Files S3 Vectors 记忆 React 框架 LLM 反思 （推理） APIs Tools/ MCP Agents 思维链 Data 上下文 业务数据基础 Bedrock Knowledge Base S3 Tables Amazon S3 S3 Vectors --- Page 17 --- 数据如何支持 Agentic AI 应用 Natural Language Query We just hired Amy Please start onboarding Orchestration Agent 任务规划 Agent StrDucattuarbeads edsata 构建定制化的入职计划 Vector search 入职伙伴 Agent Unstructured Data lakes Vectovria s tores 推荐需要联系的人 data RAG 异常处理 Agent ReaSl-ttriemame dinagta s estrrveicaems ing 立即适应变化 --- Page 18 --- Agent 记忆组件 短期 长期 当前任务或当前会话（线程） 当前用户或持久化的应用级别 Agent state Messages Semantic Profiles Episodic Prompts Action plan, Conversation Semantic User profiles Historical System Prompt, data shared history, context relevant and similar data interactions and instructions between loops, sequence of for the outcomes (RL) scratchpad Messages requested task 通常使用 通常使用工具在向量 根据应用代码、 依赖 Agentic 框架 数据存储库和企业数据库中 文件或键值存储 的后端数据存储 进行检索 进行版本控制 构建者缺乏控制 数据可访问性和质量 提示词调参 --- Page 19 --- 不只是记住， 还要智能！ Agentic memory is key 情境智能 用户偏好 知识留存 通过情境理解和模式识别提 通过记住跨多个会话的偏好 通过持久记忆和持续学习 升响应的准确性和相关性 和历史对话实现个性化交互 能力实现复杂的问题解决 --- Page 20 --- 两种场景：有记忆 vs. 无记忆 --- Page 21 --- S3 Files 像传统文件系统一样工作 Works like a traditional file system 文件和目录反映 S3 桶中的内容 • 从任何桶创建文件系统，无需数据迁移 • 智能地将数据存储在文件系统中，实现毫秒级延迟访问 • 文件系统更改立即对文件应用程序可见 • 更改在桶和文件系统之间自动同步 S3 Files --- Page 22 --- Amazon S3 Vectors 核心能力 典型适用场景 • 原生支持向量的云对象存储 • 数据湖上的语义搜索 • 向量上传、存储和查询成本降低高达 90% • 批量检索 pipelines • 零基础设施管理 • 与向量数据库形成互补架构（冷热分层） • 支持数十亿规模，并具备元数据过滤能力 • 对成本敏感的大规模向量存储 • 与 Amazon S3 相同的持久性和可用性 延迟（P95） 向量规模 QPS >100 ms 数十亿 <100s/index --- Page 23 --- 存储 — AI Agent 的外部大脑 Amazon Storage - The External Brain for AI Agents 可扩展的知识库 Amazon S3 Vectors Long term memory and RAG 持久化记忆与上下文 Amazon S3 Files & FSx 工作记忆 AI Agent Short Term Memory 多 Agent 协作 Amazon Amazon S3 S3 data lakes Real-time data Tstarbealemsing 鲁棒性与弹性 Data Foundation --- Page 24 --- S3 Files — 大规模 Agent 场景的持久化存储 For agents operating at scale • 每个 Agent 都需要相互隔 离的数据访问 • 通过 Access Points 实现数 据访问隔离的 Agent 工作 S3 Files 区，存储空间按需自动扩展 • 每个文件系统可扩展至 10,000 个 Access Points Agents run time S3 Access Points --- Page 25 --- 推理与数据管道 Inference and Data Pipelines S3 Files S3 Express One Zone S3 Vectors 大规模存储和查询向量嵌入 为 ML 提供 S3 的文件系统访问个位数毫秒级延迟 用于 RAG 工作负载 框架和推理管道 用于模型服务和实时推理 相比专用向量数据库 S3 Files 提供低延迟性能的原生 NFS 大规模并发下的一致高吞吐量 成本降低多达 90% 挂载，无需数据复制或重新格式化 适用于需要快速读取的工作负载 同时具备完整的 S3 持久性 适用于模型加载和 适用于热模型缓存和实时服务 适用于语义搜索和检索增强生成 批量推理工作负载 --- Page 26 --- RAG in Action User Input Text Generation User Prompt Large Response augmentation Language Workflow Model 文本生成 Embeddings Context model Embedding 0.89 -0.02 -0.53 0.95 0.17 -0.38 Data Ingestion Semantic search Workflow 数据摄入 S3 Vectors Embeddings Document chunks Data source model --- Page 27 --- Agent 在 EKS 上的部署 Amazon Cognito Amazon S3 Files (SessionManager) 用户认证 Session State Artifacts & Scratch 旧金山的天气 怎么样？ Amazon EKS 聊天 MCP 天气 界面 天气 Server API Alice Agent Amazon S3 Amazon S3 Vectors (Data Lakehouse) RAG LLM User Content Long Term Memory --- Page 28 --- 开始使用 - Getting Started 1 2 3 识别您的工作负载 构建您的数据基础 加速投产 训练/微调？ 以 S3 作为您的数据湖起步 使用决策框架 将 GPU 数量匹配到存储层级 进行训练存储选择 使用 FSx for Lustre 用于训练 Agentic AI？使用 Agent 记忆分类法 将数据映射到记忆类型 添加 S3 Files + S3 Vectors 将存储原语映射 （短期 vs. 长期） 用于 Agent 记忆和 RAG 到 Agent 需求 联系您的亚马逊云科技存储专家 | aws.amazon.com/fsx | aws.amazon.com/s3 --- Page 29 --- 关键要点 - Key Takeaways 训练与微调 Agentic AI 01 存储匹配集群规模 01 您的数据是差异化优势 <16 GPU: S3 | 16-1000: FSx for Lustre | >1000: S3 Express Agent 的能力取决于它们能访问和推理的知识 02 消除 I/O 瓶颈 02 Agent 需要记忆 短期（S3 Files）用于工作状态+ 长期（S3 Vectors）用于 闲置 GPU 每小时 $32 ——存储吞吐必须匹配计算 RAG 和回忆 03 从 FSx for Lustre 开始 03 专用存储原语 S3 Files、S3 Vectors、S3 Tables ——每个解决不同的 POSIX、亚毫秒延迟、1 TB/s ——覆盖大多数训练工作负载 Agent 访问模式 存储是生成式 AI 的基础——从训练到生产 Agent --- Page 30 --- Thank you ## 三、会议回放视频转录内容 Hello there This should be experience Or Nihao I should say This should be an experience This should be the first time I get to present in front of a group of people Where the presentation is entirely in Mandarin And I get to look a little script here Make sure that I'm saying the right things as we go So first and foremost Thank you Like I said, my name is Christian Smith I run the worldwide specialist team For storage and data protection Look, today we're going to talk about something today That often gets overlooked Which is storage in the conversations around GenAI You know, everybody talks about models and GPUs and frameworks And I'm actually going to show you how your storage strategy Is actually going to be the make or break factor In your GenAI performance and cost efficiency So let's start with some of the fundamental truths Look, when you're doing GenAI The real differentiator is your data It's not always the model So everyone has pretty much the same access To a set of models in their country or Region And what makes the application unique Is the quality, the relevance And the accessibility of your data And as a matter of fact, when I've been here I've been talking to a lot of customers And this has been the number one thing that comes up Is like, how is my data quality In order to feed it into my models So that's always a great starting place But whether you're building a model from scratch Or customizing one with fine tuning Or powering an AI agent It's really going to be tied to How well your storage layer delivers your compute And this isn't hypothetical The world's largest data lake today Which store all of this data Run on S3 And the question isn't whether your data Is in the cloud, it's whether your storage Architecture is optimized for the Specific demands of GenAI workloads So let's frame this up a little bit There are three distinct phases Or workloads This one has a very distinct storage requirement So first there's training This is building a foundational model from scratch We're talking trillions of tokens, petabytes of data Thousands of GPUs running in parallel Storage needs to deliver massive Sustainable throughput If you're building your own LLM There's a good chance you're probably working with Just due to the sheer quantity of LLM Sure quantity of GPUs that you're using Second, there's fine tuning Where you're taking a pre-trained model And you're applying your domain-specific data Techniques like LoRA and RLHF Or full fine tuning And the scale is smaller But the GPU utilization is still present Third is agentic And with agentic, it's about AI agents reasoning over your domain Data using tools Creating complex tasks autonomously Agentists need fundamentally different Storage patterns, persistent memory Vector search, and file-based system access Let me walk you through what storage Looks like in each of these phases Here's our portfolio we're going to talk about today There's our object storage portfolio There's Amazon S3 There's S3 Onezone Express This is really about modern architectures It's about storing data at object scale And petabytes of data Then we have FSx for Lustre Fully managed parallel file system Sub millisecond latency Sub millisecond throughput Aggregate throughput scales into the terabytes of data per second And then our newest addition So we start to talk about platforms We talk about data platform We talk about S3 vectors S3 files and mount points Which is really about storing Vector embeddings Accessing your S3 data natively As a file system, so think agents Are providing mount points To provide a file system interface For high throughput, high bandwidth environments Each of these will solve a distinct pattern That we're going to talk about today So we're going to dig into training first So this is probably the highest Throughput use case that we have Which is high throughput, high bandwidth And really this is important Because this is where economics Really matter So fine tuning Foundation models On your domain specific data Really storage becomes the Bottle neck in almost all the cases That we have seen today And it's because of You want to keep your GPUs As busy as possible Because if they're not busy That's wasted dollars So for example This is from a while ago Like a p4, p5 instance Is about $32 per hour And so every time that GPU Is not fully utilized That's money wasted And storage again Is the bottle neck here But when you choose Starve GPUs Let's dive in a little bit here So let's put the guide together So the first one is Let's Amazon S3 So architecting for scale So when you're training Smaller models and fine tuning In the kind of 1 to 16 instance S3 is your straightforward choice So you get streaming throughput Virtually unlimited capacity And you only pay for what you use S3 integrates natively With SageMaker and PyTorch data S3 和 贴文件 这项目的功能是 用 ML 和 自行管理的 ML 和 工程的 功能 这项目的功能是 用 LoRA、QLoRA S3 和 单码的 训练 这些功能不需要 需要 训练的 技术 S3 会很难 如果 训练 技术 或 训练 是 功能需要 百分百的 技术 如果您有这些用途的情况，请请我们一起工作 我们有一个很好的团队的市场考验 可以建立最好的图案 Lustre，我们再次提到 Lustre 它会在16-2000个时期的时间中 提供的能量是 parallel IO 有千万的客人，在一秒钟的时间中 保持 GPUs 永远地浮现 调整一秒至一秒的空间 它是 natively with PyTorch deep speed, Megatron, no code changes 一个重要的因素是 我们有了一些能量的功能 你可以连接到 S3 的预算 然后将资料放在资料中 这意味着你不需要 预算很多的资料 你只需要预算大量的资料 在你的热工程中 You get things like LZ4 compression to reduce storage And it has things like EFA So EFA is elastic fabric adapter That's our high speed low latency network You get NVIDIA GPU direct This equates to getting 12 to 1300 GB per second per instance The big operational planning here If you get into the billions of files Work with us Because we're going to be in this Working with you to make sure that data gets distributed Across metadata servers For the best throughput and performance Now when you move beyond 1000 GPUs Think P5s or Trainiums This is where we talk about mega scale So this is where S3 Onezone becomes your preferred choice It delivers 10x lower latency than standard S3 It's consistently high throughput at high concurrency It has request level parallelism It scales linearly The directory bucket structure is optimized for ML access patterns This is for customers that are strongly cloud native And they prefer an object interface You can put mount points You can put a file interface in front of S3 Express Onezone But if you're at this scale You're going to be building your own custom data loaders And you're going to want to be working with us To get the highest throughput possible In practice Almost all of our customers have started out with Lustre This has been the preferred platform Preferred platform For all of our customers doing training and fine tuning It's simple to get started It works out of the box It natively integrates with S3 You get linear scaling You have a choice between persistent and scratch deployments To match durability and workload lifecycle My recommendations start sizing with throughput requirements So this is an important point So calculate how many GPUs you're going to need Or you're going to get to Size it for throughput first Capacity second Because this will make your life very easy When you continue to add more and more data And don't forget to use intelligent tiering FSX intelligent tiering For your underlying S3 data optimized cost Without sacrificing performance Adobe great example They launched the Firefly generative AI models In just nine months using FSx for Lustre To power their high performance model training When you're competing in the GenAI market Time to model is everything And storage throughput directly determines that timeline We built a decision framework here So just recapping everything But I said at the beginning If you have under 16 Just use S3 Because that's most likely where your data is at Realistically, if your data is in EFS Or EBS or FSXZ or N Just work from there Like don't spend the time moving your data around Or shuffling your data around Just to make it available to GPU frameworks If you're in the 16 to 500 This is most likely going to be Lustre If you start to get between 500 to 1000 You start to get in the place Where you can make a decision of going between Lustre Or S3 One Zone Express And it really depends on Are you going to be writing your own data loaders or not And then of course you get bigger than that You're going to want to go with S3 One Zone Express It has the best economics at that scale Write this down Take a picture It'll save you a lot of headaches later Okay, let's shift gears completely And talk about agentic AI This is one of the fastest growing areas In our organization And it requires a fundamentally different Storage architecture I'm going to walk you through a little bit Of a progression here About where we've come from And where we're going to And why these storage classes matter As we get to agentic I don't know if we know everything yet But these are the patterns that we're seeing emerge So first Look, in our life This has been the journey we've been on I can't read that But I think it says something along the lines of We started out with kind of classic machine learning We focused on analytical tasks Pre-defined rules We then got to generative AI assistants Think chatbots All those websites you went to And you asked it a question And it didn't answer you And then it directed you to customer service That we thought were really cool At the time, but they were really useless Well, now we're getting to fully Agentic systems that are emerging Autonomous workflows Where multiple agents are collaborating And really with minimal human oversight That's where we're heading to As we move into the spectrum The demands of your data foundation increase exponentially You know, agents need far more data And in far more ways Than this like simple chatbot ever did So this is the big picture That kind of plays out what we knew today You know, this is going to be start with your data foundation Agent need to understand And consume structured data Like financial reports Customer records There's unstructured data like images Videos, documents This is really your data that sits in your lake house Right? This is your foundation of all your data For RAG, agents need vector search And knowledge bases To access relevant content S3 vectors and Bedrock Knowledge base serve this role For agentic memory And this is critical Agent need short term memory For storing intermediate content Like task state Or enabling multiple agents To exchange information S3 files and S3 express Are really going to be well suited here Long term memory Allows agents to recall information Across sessions, users And be more personalized over time So stored as semantic embeddings In S3 vectors And finally here hosting the LLM models That give agents their reasoning capacity This requires fast model loading over here And context retrieval All of this is built on Things like Bedrock, AgentCore, SageMaker But a lot of customers are doing this themselves In EKS with their own self managed models So they're building these layers out And using these primitives on their own Let me give you a real concrete example here So this is an example of an Agentic orchestration Task planning So this is let's say we just hired Amy Amy is going to be onboarding soon A task planning agent right here Is going to build a tailored plan Via API calls using Amy's profile To assign task provision tools Schedule activities Pulling from structured databases An onboarding buddy curates a personal learning path Using vector search across handbooks and wikis And recommends coworkers based on roles and interests And at the bottom you got an exception handling agent Monitoring for disruptions Like a delayed laptop or a changing start date And adapts in real time feeding this state back up Into the other layers Each agent needs different data Different access patterns and different storage services To function effectively So I'm leading to memory And this is the path that we're on Because this is the new area Agent in order to get autonomous In order to get agentic Are going to need more and more memory And memory of different types And the way we look at it is two types of memory There's short term memory Which is your current task or session This is going to store your agent state It's going to store in-flight tasks It's going to store Messages Checkpoints for crash recovery It's tightly integrated with frameworks Like LangChain, CrewAI, Strands And typically uses high performance shared storage Like S3 Express, S3 files, or instance memory Really, you're thinking minutes of retention here But most people don't delete this Becomes a good source to refine later And create new contextual state Long term memory Is really about persistence across users and sessions This includes semantic content User profile, Historic interactions So every time I'm my agent And I have the Christian bot that I use My Christian bot has Prompts That tells That I've gone through all my documents It has all my writing style It has all my Slack Messages That I've ever written It knows who I am So my agent has this long term memory That every time it invokes It knows who I am And how I write and how I respond And it's pulling it from our long term memory So that it has this Prompt That says Write like Christian Or respond How would Christian respond? Typically condensed from short term memory So we're going from short term memory Down to long term memory It's really embedded It's stored in vector databases At large scale and retrieval This is where S3 vectors Which is purpose built for this The last category is Prompts It's a form of long term memory Stored in like key value stores Like DynamoDB or ElastiCache You can use it to reduce repeated LLM calls Like in caching Prompts To reduce agent operational costs So agentic memory is the key Look proper agentic memory Unlocks three transformative capabilities I keep stepping in front If you want to take pictures I'll move out of the way Contextual intelligence here Where the agents not just only understand What you're asking But why you're asking it It's like pattern recognition Across interactions There's user preferences Which is about personalized interactions The agent remembers how you work How you communicate across sessions And then there's knowledge retention Which is continuous learning The agent gets smarter With every interaction So it's basically fine tuning itself All as it goes These aren't really nice to have They're essential for enterprise adoption Without memory, agents can't maintain Conversational continuity Can't adapt from feedback And can't develop user-specific understanding Here's an example With memory and without memory So without memory, you're asking questions And you're always going to start out The same place every time you start It's going to always start out With something like who are you And it's going to go Continue to ask those questions From there It's going to have no sense Of your previous interaction It's always Think of it like a squirrel Or like a gerbil Where it's a goldfish Which has a memory of two seconds Like it's always excited to see you It's always going to say the same thing It's always going to respond the same way Now you start to get it into memory And this is where you start To get that repeated interaction over time So as you get into memory You can interact with things That remembers your contextual History It remembers how it responded to you It remembers how you responded to it So that when you come in the next time It's going to have contextual awareness Of who you are, what you do And what types of questions that you're asking So I've given you the history of memory I think we've led up to that Which have had nothing to do with storage yet So let's dive into this first So when we talk about this type of agent memory We talk about really kind of two things here We're going to first talk about If we're using S3 as our data foundation And it has all of our Organizational line of business Enterprise type data S3 files is a great way To connect that data to your agents So your agents can directly interact With that data through a file system interface And almost all agents are built Expecting a file system interface It's the way they operate Now from an agent perspective Agent are going to be creating code They're going to be executing code They're going to be participating They're going to be creating contextual memory Via text files, MD files They're going to be storing long term memory In MD files or storing short term memory In MD files And so this is a great place That you can then also interact with that data But store that short term memory state Directly to a file system And then over time That data gets synced directly into your S3 bucket So that you have this long term Agent state over time When you deploy like this There's no custom code There's no manual sync jobs It just works So an example of this is We were talking to a bioinformatics company They have genomics data They have agents start to parse Through this genomics data And each agent has a task Like you're looking for this type of data When you find this type of data You're going to invoke this other agent Who's going to go search for this type of cancer Or tumor in that data As it's going along It's building up code To go find that data It's feeding that data into LLMs It's persisting memory and state So that the next time you interact with it Or tell it to do something different It knows where it started It hands off that data To another agent Which is going to compare it against a known database Of existing tumors and cancers Looking for that genomic defect And these things will run over and over and over again Data is in S3 Agent accessing through a file system interface Very easy to connect these two things together And persist state No manual custom code No jobs, it just works Now, I don't know Most people here know what S3 vectors are So S3 vectors So we've had S3 As an object store for years What you're seeing is As a data platform It's moving up into the value chain And here's what's happening here We notice that customers are running A lot of things like OpenSearch Or their vector databases A lot of these vector databases Get very costly to run In the billions of rows They're very expensive because they treat all data as hot And they have to run on kind of like The most expensive hardware that's out there What we said was You know, the semantics of puts and gets are sort of similar We could add a k-top to it We could add a Similarity index to it And we can create a vector database In S3 through a vector bucket The result is You can store billions of vectors Very inexpensively It's like 90% lower cost The tradeoff is you're going to have a higher query time And so if you're dealing with things like agents Does that matter? If it's 100 milliseconds versus 1 millisecond Does that matter if you're saving 90% on cost? Most of the time, no Really where that High performance rag database It needs to come in Is when you're doing something interactive Like with a website Where there's a time to element For customer experience So again, very Advantages for people that need multi-billion dollar Multi-billion row vectors It does work with ElastiCache OpenSearch so you can tier vectors into it But it's really about How do I take that memory Long term memory state And store it in something cheap and inexpensive That provides that performance Necessary to meet the requirements And so really This is the layering of what it looks like So how do we create storage Is the brain for agents It starts with long term memory and rag And using S3 vectors You go into short term memory And it's S3 files You can of course use FSX Or EFS in here too It's also for working memory too So if you're just doing things Like working with data That you have stored in your bucket It's robustness and reliability here With things like The 11 nines of S3s Durability And you can do multi-agent collaboration We really recommend building Your data foundation on S3 Because of the resilience, cost effectiveness And integration with both 亚马逊云科技 analytic services And the broad ecosystem Of third party services out there I'm coming to a close here soon So we're going to cover Two more use cases One, we've seen a lot of people Deploy open claw at scale And they're thinking about Doing it as a service So they're going to provide it As a service to their end users We do think that open claw Will become kind of the foundation That a lot of ISVs Will build their base off of And turn it into a product When you start to do that And you're turning your Open claw environment Into a SaaS experience There's something that you want to think about Which is Okay, I need a file interface I need the ability to store that data Eventually into S3 For long-term durability And cost effectiveness But I also need Tenant isolation here So I can't have Customer X Interfering with customer Y Their data can't cross or commingle So we give you a very cost-effective way To create tenant isolation You can create up to 10,000 access points For file system Each with its own isolated security model You can almost think about it Like giving each agent its own sandbox You can also do this Inside your organization too So if you're building A tenanted agent model For your organization This is a great way to do it Last thing Storage for inference And data pipelines So storage for inference Specifically Self-manage on EKS It's really about three different services By the way, if you go on Our container on the Amazon I can't say 亚马逊云科技 亚马逊云科技 And you go to the EKS Landing page There is a link To go to workshops Our top workshop right now Is running inference On EKS And running fine-tuning on EKS These are free You can go take it anytime you want You'll get like a hands-on lab through it So walking through this So inference Especially on self-manage EKS S3 files or EFS For model loading and batch inference The native NFS provides low latency No copying of data No reformatting needed S3 express one zone Is really about hot model caches So if I'm deploying thousands of EKS nodes And I change the library That load time to reactivate All those containers becomes really critical So think of these as low capacity But really high throughput So I may only have a couple gigs But I'm going to burst at times Up to 200 to 300 gigabytes per second As I flip my model This is a great service to do that with here And then S3 vectors for your rag pipeline Really grounding your foundation models And providing like the 90% lower cost Over specialized vector databases This is rag in action So here's how it works On the bottom here You're going to be taking your data source Turning it into document chunks You're going to be running an embedding model in it You're going to be storing those embeddings In S3 vectors At query time User input goes through It's going to query the same embedding model And the vector database looking for semantic similarity It retrieves the most relevant passages back up And passes them to the context LLM Vectors and semantic search The foundation for rag S3 vectors Makes this really simple and scalable and cost effective In cases where kind of that access time Can be built into what you're doing Which tends to be a lot of time in Prompt documentation And here is the state for agent deployment and EKS When you put it all together When I talk about EKS workshops And doing self-manage This is how we think about putting it all together So you have here An agent-maintaining state In S3 accessing long term memory In S3 vectors They're storing user content In the data lake And connects to external servers Like MCP servers And it can use authentication through Cognito So even in a simple like agent example here You can see how storage services kind of map Into each distinct phase of this Short term memory, long term memory rag Foundational data And for frameworks if you're using things like strands We have S3 session manager Can provide native session state out of the box How to get started Identify your workload For training, match GPU count to storage Map your data to the data types Memory types Build your data foundation Start with S3 Accelerate to production Use the decision framework for what we showed earlier For storage use agent memory taxonomy To map storage, primitives to agent And please engage your specialist if this gets hard I think you're giving me a time And I don't know what that says But I think it's a minute Key takeaways Again fine tuning Agentic AI Great picture to take here I won't rehash it I think we've covered it enough But your data is the differentiator here And so agents are only going to be as good As the data you provided That's it Thank you I had a great time