2026 年云栖大会，Hologres 发布全新 5.0 版本，以"Agent Native 的全模态实时数仓"为核心理念，提出"从实时数据，到可靠行动"的演进方向。Hologres 负责人姜伟华在演讲中指出，随着 AI Agent 成为数据系统的全新用户角色，数据基础设施正面临从"服务人和报表"到"服务 Agent 自主决策与行动"的范式转变。Hologres 5.0 围绕实时完整的数据、可理解的语义、可信可控的行动、自适应执行四大支柱进行全面升级，并首次发布数据沙箱、图检索、Semantic View、User Defined Agent、OpenMemory、HoloScope 等多项重磅能力，致力于成为 Agent 时代企业最可靠的数据底座。
阿里云智能集团 · 计算平台事业部 Hologres 负责人 姜伟华
从服务人到服务 Agent：OLAP 迎来全新挑战
过去，数据系统主要服务于业务人员和 BI 报表，用户通过 SQL 提问、人工解释、人推动下一步行动；而现在，Agent 正在成为新的"数据用户"——它持续运行、自主组合上下文、检索推理并调用工具，结果直接回流到下一轮判断与行动中。
Agent 带来的价值显而易见：更低门槛的自然语言理解复杂数据、持续实时感知变化、从洞察走向闭环行动。但与此同时，它也给数据系统提出了四项全新挑战。
第一，实时完整的数据。 Agent 需要同时处理结构化、多模态、关联关系等多种数据形态，且要求数据实时可查。过去数据可能散落在各个孤立系统中，但现在 Agent 需要访问完整的数据视图，才能做出足够深入的洞察。
第二，可理解的语义。 指标口径、实体关系等业务语义必须内建于数据系统，而非依赖人工反复解释物理表结构。“什么是有效订单”——这类过去只存在于 SQL 和业务人员脑子里的概念，现在必须被显式表达出来，让 Agent 能够理解。这也是为什么近期与 Agent 相关的本体论研究变得非常热门。
第三，可信可控的行动。 Agent 的操作必须有权限围栏、隔离环境和完整观测，确保每一步行动可审阅、可追溯。OLAP 系统往往直接服务于 2B 或 2C 的线上系统，Agent 如果随意操作，很容易影响线上稳定性。即使 Agent 是善意的，一条未经优化的 SQL 也可能瞬间耗尽实例资源，导致系统崩溃。
第四，查询不可预测，功能优先于性能。 Agent 的访问模式和资源消耗难以提前预判，数据引擎必须具备自适应优化能力。Agent 更关注功能是否正确，而不会像人类工程师那样精心调优 SQL 性能，但 OLAP 系统对查询性能极其敏感，如何让 Agent 生成的 SQL 稳定高效地运行，是一个巨大的挑战。
数据沙箱：让 Agent 拥有安全的行动空间
Hologres 5.0 首次发布数据沙箱能力。 这是基于当前生产实例创建的隔离Agent工作空间，支持Agent任意增删查改，而不会影响原实例的生产数据。
数据沙箱的设计理念是在生产实例之上创建一个可写、可验证的隔离环境。Agent 可以在沙箱中自由执行分析、修改指标口径、实验表结构优化，所有操作均使用独立计算资源，不影响生产系统。验证通过并审批后，变更才合并回原实例。
这一能力对 Agent 场景尤为关键。在过去，Agent 直接操作生产数据存在巨大风险；有了数据沙箱，Agent 获得了一个"自由但安全"的行动空间，可以先实验、再验证、最后交付，真正实现从洞察到执行的闭环。即使面对 PB 级数据，Hologres 也能快速克隆出沙箱环境，让 Agent 在其中执行长时间、多步骤的复杂操作。 Hologres Agent Demo 演示 >>
全新图检索：从"找相似"走向"懂关系"
在标量查询、全文检索、向量检索的基础上，Hologres 5.0 全新发布图检索能力，基于 AGE 扩展，通过 SQL 调用 Cypher 子集并与普通 SQL 组合分析。这标志着 Hologres 从"分析与相似检索"迈向"实体关系与路径查询"，成为真正的全模态数仓。
图检索的核心价值在于回答"为什么相关、如何关联"。过去计算数据更多是找各种数据之间的相关性，向量检索就是一种相关性的检索；但在本体论的时代，关系、知识图谱变得非常重要。以电商场景为例，传统向量检索可以找到相似商品，但图检索可以沿着"客户→下单→订单→包含→商品→规则命中→渠道→品牌"的关系路径进行多跳查询，支撑知识图谱构建与 Agent 决策。
在 Agent 应用中，图检索的价值尤为突出：Agent 从一个业务问题出发，通过图检索找到正确的数据、理解指标口径、生成有据可查的分析结果，让 AI 推理不再"凭空猜测"，而是沿着真实的关系网络展开。姜伟华透露，目前已有多个内部和外部客户基于 Hologres 的图能力构建知识图谱。
Semantic View：让人、BI 和 AI 使用同一套业务语言
同一指标多套算法、星型模型 JOIN 复杂、一对多 JOIN 造成指标放大、AI 只看物理表容易猜口径——这些长期困扰数据团队的痛点，在 Hologres 5.0 中通过 Semantic View 得到系统性解决。
Semantic View 允许用户以声明式方式定义表间关系（TABLES / RELATIONSHIPS）、业务维度（DIMENSIONS）和指标（METRICS），由引擎自动完成最小 JOIN 子树计算、表裁剪与先聚合后关联，规避 fan-out 问题，保证同一条记录对同一维度至多贡献一次。最终生成 YAML 语义模型，供 BI 工具和 AI Agent / NL2SQL 直接读取。Semantic View 兼容 Apache Ossie，这是当前最流行的语义层定义标准之一。
这意味着，业务人员、BI 报表和 AI Agent 终于可以使用同一套业务语言进行数据交互，不再需要反复解释物理 JOIN 与指标口径。对 Agent 而言，Semantic View 让它能够"听懂"业务语义，大幅提升数据访问的准确性和效率。
User Defined Agent：让 Agent 与数据深度绑定
Hologres 5.0 发布了 User Defined Agent 能力，开发者可以用 Python 快速定义一个 Agent，支持 Memory 和 Tool 调用，并通过 DataFrame API 实现批量调用与增量运行。
以一个实时评分场景为例：用户有大量文本形式的商品评价数据，希望将这些评价自动转换为好评分数。首先用 Python 简单定义一个 Agent 类，声明模型连接、Prompt 和工具集；然后通过 DataFrame 的 apply_agent 方法逐行触发 Agent 事件循环进行批量分析，可以直接运行查看结果是否符合预期；符合预期后，将同一 DataFrame 物化为 Dynamic Table，以增量刷新方式持续运行，仅处理新增行。这样就了构建一个一分钟延迟的实时数据管线。整个过程通过 Agent 与 Dynamic Table 的深度结合实现：Agent 不再是外挂式的应用，而是与 Hologres 数据深度融合的"原生居民"，可以直接在数据仓库内运行，利用 Dynamic Table 的增量加工能力持续处理新数据，真正实现"数据驱动行动"。
这一能力本质上是将过去的用户定义函数（UDF）升级为用户定义 Agent。Agent 可以直接与 DataFrame 联用，就像一种 AI 算子，让传统的数据处理流程具备了智能分析能力。
OpenMemory 与 HoloScope：服务所有通用 Agent
Hologres 5.0 还发布了两项面向通用 Agent 生态的基础设施能力：OpenMemory 和 HoloScope。
OpenMemory 是通用 Agent 的记忆层，类似 Mem0，由 Hologres 提供存储底座。它支持跨会话、跨设备的记忆协作，实现"感知→记忆→行动"的完整闭环。在评估记忆能力最权威的基准测试之一 LOCOMO Benchmark 中，OpenMemory 以 96.82 分的成绩达到世界最高水平。 姜伟华透露，相比没有记忆的基线，OpenMemory 能够节省 90% 以上的 Token 消耗。OpenMemory 的接口基本兼容 Mem0 方案，面向所有通用 Agent，并非 Hologres 专用。
HoloScope 则是通用 Agent 的可观测平台，同样由 Hologres 提供底座。当 Agent 真正放到生产环境后，需要一套完整的可观测、验证、CI/CD 体系来保证稳定性。HoloScope 支持多模态追踪、100+ 指标接入、全量 Session 回放与评估，以及对比和 A/B 实验能力。结构化 Trace 和 Session 回放让 Agent 的每一步行动都可追踪、可审计。
四大支柱：构建 Agent 时代的完整数据底座
姜伟华在演讲最后将 Hologres 5.0 的能力总结为四大支柱：
在数据层面， Hologres 实现统一访问内仓、数据湖、MaxCompute 和 OSS，支持结构化、半结构化（JSON、Variant）和非结构化数据的全模态存储，并通过 Dynamic Table 进行增量近实时加工，生成更适合 Agent 访问的数据形态。姜伟华特别强调，数据分层对 Agent 尤为重要——本质上是数据的清洗和加工，让数据口径更加对齐，Agent 就能访问到更好的数据。今年 Dynamic Table 还新增了级联加工能力，上游第一级加工后，下游会逐级触发，避免了以往多级加工延迟累加的问题。
在语义层面， Semantic View 和 Hologres Agent 共同描述业务语义与指标口径，配合图检索能力，让 Agent 不仅能找到数据，还能理解数据之间的关系。从传统的查询"找相似"，升级为"找关系、找路径"。
在行动层面， 完整的权限体系（包括新增的行级权限）为 Agent 明确可操作边界；Skills 默认 Readonly 加 Serverless 执行，不影响生产实例；数据沙箱提供Agent自由安全的行动空间；User Defined Agent 让 Agent 与数据紧密结合。
在执行层面， Adaptive Execution（分阶段自适应执行）和 History Based Optimization（HBO，基于历史的优化）让查询无需精细调优就能稳定执行；Serverless Computing 让大 Query 稳定执行；Liquid Table 让表按需 Resharding，弹性应对不可预测的 Agent 访问模式。
完整数据、可理解语义、可控行动、自适应执行——这四者的结合，让 Hologres 5.0 真正成为 Agent Native 的全模态实时数仓。 Hologres 的目标是让数据不仅仅是实时数据，更让 Agent 在实时数据上能够可靠执行。
性能表现：多项业界第一
即使在 Agent 时代，性能依然至关重要。姜伟华在演讲中分享了 Hologres 今年的几项亮眼成绩：在向量检索领域，VectorDB Bench 测试中达到 9.75 万 QPS，位居业界第一；TPC-H 3TB 测试成绩在 TPC 官网上同样是业界第一；TPC-DS 10TB 已向 TPC 官网提交成绩，正在审核中，代表三倍于目前榜首的性能水平。
这些成绩背后，是 Hologres 在存储引擎和查询优化上的持续投入。今年 Hologres 提供了Liquid Table无损 Resharding 能力，解决了业界长期面临的难题——当数据量增长时，如何在不影响性能的前提下重新分布数据。同时，Dynamic Table 持续增强与 AI Function 的整合，让增量近实时加工能力更加强大。
结语
Hologres 5.0 的发布，标志着 Hologres 从"多模态分析平台"向"Agent 原生数据底座"的战略升级。在 AI Agent 快速发展的今天，数据系统不再只是被动的存储和查询工具，而是主动参与决策和行动的核心基础设施。从实时数据到可靠行动，Hologres 正在为 Agent 时代构建最坚实的数据底座。
正如姜伟华在演讲结尾所说：“简单来说，Hologres 5.0 的产品口号就是——从全模态实时数仓，变成 Agent Native 的全模态实时数仓。让数据不仅仅是实时数据，我们希望让 Agent 在实时数据上能够做可靠的执行。”
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In 2026, the Cloud Habitat Congress, Hologes publish, a new version of 5.0, with the idea of "Agent Native's full-state real-time silo" as its core, proposes the direction of "from real-time data to reliable action". In his speech, Kang Weihua, head of Hologres, noted that as AI Agent became a new user role for data systems, the data infrastructure was facing a paradigm shift from "server and report" to "service Agent autonomous decision-making and action". Holloges 5.0 is a comprehensive upgrading around real-time complete data, understandable semantics, credible and manageable actions, self-adapted implementation of the four pillars, and the first publish data sandbox, chart search, Semantic View, User Defined Agent, OpenMemory, Holoscap, among others, are committed to becoming the most reliable data base for Agent-era enterprises.
Alibaba Cloud Intelligent Group.
Agent:OLAP brings new challenges
In the past, the data system used to mainly serve business people and BI statements, and users used to ask questions through SQL, explain manually, and promote next step actions; and now Agent is becoming a new "data user" — it is running on an ongoing basis, grouping context, retrieving reasoning and calling tools, and the results flow directly back to the next round of judgement and action.
The value of Agent is clear: lower-threshold natural languages understand complex data, sustain real-time sensory change, move from insight to closure. At the same time, however, it presents four full new challenges to data systems.
First, complete data in real time. Agent needs to deal with multiple data patterns, such as structured, multi-modular, associated, and requires data to be available in real time. While data may have been scattered across isolated systems in the past, now Agent needs access to complete data views in order to be able to make a sufficiently deep insight.
Secondly, understandable semantics. Operational semantics such as indicator calibres, physical relationships, etc. must be built-in in data systems rather than relying on manual reinterpretation of physical table structures. “What is a valid order” — concepts that used to exist only in SQL and in the minds of operators must now be expressed in a visible way so that Agent can understand. And that's why the introspective research related to Agent has recently become very popular.
Third, credible and manageable action. Agent must operate with a permission fence, a quarantine environment and complete observation to ensure that every step of action is accessible and traceable. ORAP systems tend to directly serve 2B or 2C on-line systems, and Agent can easily affect online stability if it operates at will. Even if Agent was kind, an unoptimised SQL could quickly exhaust the resources of the case, causing the system to collapse.
Fourthly, queries are unpredictable and functions take precedence over performance. Agent ' s access patterns and resource consumption are difficult to predict in advance, and data engines must have the capability to adapt to optimization. Agent is more concerned about the correctness of the function, rather than fine-tuning SQL performances like human engineers, but the OLP system is extremely sensitive to query performance, and the challenge is how to make SQLs generated by Agent work in a stable and efficient manner.
Data Sandbox: Give Agent a secure operating space
Hologres 5.0 First publish data sandbox capability. This is the segregated Agent workspace created on the basis of current examples of production and supports Agent ' s arbitrary additions and deletions without affecting the production data of the original example.
The design concept of data sandboxes is to create a writeable and verifiable segregation environment based on production examples. Agent is free to carry out analysis, modify the calibration of indicators, optimize the structure of the test sheet in sandboxes, using independent computing resources for all operations without affecting the production system. Once the validation is approved and approved, the change is consolidated back into the original example.
This ability is particularly critical for the Agent scene. In the past, Agent has had a great risk of directly operating production data; with the data sandbox, Agent has been given a "free but safe" space to experiment, validate, deliver, and truly achieve a closed loop from insight to implementation. Even with PB-level data, Hologres quickly cloned the sandbox environment in which Agent carries out complex operations that are lengthy and multi-step. Hollogres Age Demo Presentation >
New Map Retrieval: From "Find Like" to "Know About."
Based on the cursor search, full text search, vector search, Hollogres 5.0 complete new publish graph search capability, based on the AGE extension, calls the Cyprus subset through SQL and combines analysis with normal SQL. It's a sign that Hologres is moving from "analysing and similar searches" to "substantive relationships and path queries" to a real full-state digital warehouse.
The core value of chart search lies in the answer to "why is it relevant, how is it related." In the past, the calculation of data was more of a search for correlations between data, and vector retrieval was a retributive search; in the age of introspection, however, relationships, knowledge maps became important. For example, in the case of electronics, traditional vector search can find similar goods, but map search can support decision-making on knowledge mapping by multi-jumping along the path of the relationship between "Customs placing a single order and "Customs Zero" and "Customs Zeros".
In Agent's application, the value of graph retrieval is particularly strong: Agent, starting with a business problem, finds the right data through a map search, understands the calibration of indicators, produces documented analysis, and allows AI reasoning to move away from "fiscal speculation" and to proceed along the real network of relationships. Kang Weihua revealed that there are now multiple internal and external clients who build knowledge maps based on Hologres ' abilities.
Semantic View: Make people, BI and AI use the same business languages
Multiple algorithms for the same indicator, star model JOIN complex, multiple pairs of JOIN magnify the indicators, AI is easy to guess from a physical table - the pains of these long-plained data teams are systematically addressed through Semantic View in Hologes 5.0.
Semantic View allows users to define inter-table relationships (TABLES / RELATIONSHIPS), operational dimensions (DIMENSIONS) and indicators (METRICS) in a declarative manner, with the engine automatically completing the smallest JoIN sub-tree calculation, tailoring and association, avoiding fan-out problems and ensuring that the same record contributes to the same dimension at most once. The YAML semantic model is eventually generated for direct reading by the BI tool and AI Agent/ NL2SQL. Semantic View compatible with Apache Ossie, which is one of the most popular semantic definitions.
This means that operatives, BI statements and AI Agent can finally interact with data in the same business language and no longer have to explain the physical JOIN and indicator calibres repeatedly. For Agent, Semantic View allows it to "understand" business semantics and significantly improves the accuracy and efficiency of data access.
User Defined Agent: binding Agent to data depth
Hologres 5.0 publish has the User Defined Agent capability, which can be quickly defined by developers using Python to support Memoory and Tool calls and to enable bulk calls and incremental movements through DataFrame API.
Using a real-time rating scenario, for example, users have a large number of commodity evaluation data in text form that they wish to automatically convert into good ratings. First, you simply define an Agent class with Python, declaring model connections, Prompt and tool sets; then you trigger the Agent event cycle by row through the application agent method of DataFrame, which allows direct running to see if the results are in line; and then you convert the same DataFrame into Dynamic Table, which is running continuously in an incremental refreshing manner, and only handles new lines. This leads to the construction of a one minute delayed real-time data pipeline. The whole process is achieved through the deep integration of Agent with Dynamic Table: Agent is no longer an outside application, but a "native resident" that is integrated with Hologres data depth, and can operate directly in the data warehouse, using the incremental processing capacity of Dynamic Table to process new data on a continuous basis, and actually achieve "data drive action".
This capability essentially upgrades the former user definition function (UDF) to the user definition Agent. Agent can be used directly with DataFrame, like an AI algorithm that gives traditional data processing skills.
OpenMemoory with Holoscope: Service All CommonAgent
Hologres 5.0 also publish has two infrastructure capabilities for universal Agent ecology: OpenMemory and Holoscope.
OpenMemory is a universal Agent memory layer, similar to Mem0 and is provided by Hologres. It supports cross-conference, cross-device memory collaboration to achieve the complete closure of the Sensory Memory Operation. In one of the most authoritative benchmark tests to assess memory capacity, LOCOMO Benchmark, OpenMemory reached the highest level in the world at 96.82 points. Kang Weihua revealed that OpenMemory could save more than 90 per cent of Token consumption compared to a baseline without memory. OpenMemory's interface is generally compatible with the Mem0 programme and is open to all generic Agents, not exclusively for Hologres.
HoloScope is a universal Agent observable platform, also provided by Hologres. When Agent is actually placed in a production environment, a complete CI/CD system is needed to ensure stability. HoloScope supports multi-modular tracking, 100+ indicator access, full Ssession playback and assessment, and comparison and A/B experimental capabilities. Structured Trace and Session Replay enabled every step of Agent to be tracked and audited.
Four pillars: building the complete data base of the Age of Agent
At the end of his presentation, Kang Weihua summarized the capabilities of Holloges 5.0 as four pillars:
At the data level, Hologres achieves integrated access to the interior, data lakes, MaxCompute and OSS, supports the full-modular storage of structured, semi-structured (JSON, Varant) and non-structured data, and produces data patterns more suitable for Agent access through incremental real-time processing through Dynamic Table. Kang Weihua stressed in particular that the data hierarchy is particularly important for Agent-- I'm sorry. Essentially, it's data cleansing and processing, so that the data can be more calibrated and Agent can access better data. This year, Dynamic Table also added cascade processing capacity, which, after processing upstream, triggers a step-by-step downstream process, avoiding the problem of past multi-stage processing delays.
At the semantic level, Semantic View and Hologres Agent jointly describe the business semantics and the calibration of indicators, with the ability to retrieve them, so that Agent can not only find the data but also understand the relationship between them. From the traditional queries, "Find similar," to "Find relationships, find paths."
At the operational level, the complete permission system (including the new line level permission) is Agent's clearly operational boundary; Skills defaults Really plus Serverles implementation, without prejudice to production examples; Data Sandbox provides Agent's free and secure operating space; User Defined Agent allows Agent to be closely associated with data.
At the implementation level, the Adaptive Exchange (phased self-adaptation) and History Based Optimization (HBO, based on historical optimization) allow queries to proceed steadily without precision; Servers Computing allows large Query to proceed steadily; Liquid Table allows tables to respond flexibly to unpredictable Agent modes of access as needed.
Complete data, comprehensible semantics, controlled actions, self-adaptation - the combination of these four makes Hologes 5.0 truly the Agent Native's real-time silos. The goal of Hologres is to make the data more real-time than real-time, and to make it more reliable on real-time data.
Performance performance: multiple industries first
Even in the Age of Agent, performance remains crucial. In his speech, Kang Weihua shared several bright-eyed achievements of the year: in the field of vector retrieval, the VectorDB Bench test reached 975 000 QPS, the first in the industry; the TPC online test results for TCC-H 3TB were also the first in the industry; the TPC-DS 10TB has submitted its scores to the TPC official network, which is under review, representing three times the current top performance level.
Behind these scores are the continuing investment of Hologres in the optimization of storage engines and queries. This year, Hologres provided the capacity of Liquid Table to do nothing to resolve the long-standing problem facing industry — how to redistribute data without compromising performance when data volume increases. Meanwhile, Dynamic Table continuously enhances integration with AIFunction to make incremental near real-time processing more powerful.
Concluding remarks
Hologres 5.0, publish, marks the strategic upgrading of Hologres from "Multimodal Analysis Platform" to "Agent Raw Data Base". Today, in the rapid development of AI Agent, data systems are no longer merely passive storage and query tools, but are the core infrastructure for active decision-making and action. From real-time data to reliable action, Holloges is building the strongest data base for the Age Age.
As Kang Weihua said at the end of his speech: "In simple words, the product slogan of Hologres 5.0 is -- from full-modular real-time to full-time mode real-time. Let's make data more than real-time data, and we want Agent to be able to do it on real-time data."
Auto-fetched and translated by AI Atlas. All rights belong to the original author.