80
Elastic open-sourced Atlas, based on Elasticsearch to implement episodic, semantic, and procedural memory for intelligent agents, achieving long-term memory storage and retrieval via MCP integration and hybrid retrieval. 
Yesterday 1378 words (about 6 min) View Source →
Sign in to highlight text and take notes as you read. Sign in now
InfoQ 2026-07-11 11:37 北京
Elastic 开源了 Atlas。这是一个基于 Elasticsearch 构建的系统,可以为智能代理维护三类记忆。
作者 | Anthony Alford
译者 | 平川
Elastic 开源了 Atlas。这是一个基于 Elasticsearch 构建的系统,可以为智能代理维护三类记忆。Atlas 通过 MCP 与智能代理集成,并保持用户记忆的相互隔离。在问答能力评估中,其 Recall@10 得分为 0.89。
Atlas 旨在解决以下问题:在面对与智能代理有长期交互历史的用户时,应该如何识别适当的上下文数据并添加到智能代理的 LLM 提示词中。按照 Elastic 的说法,加载完整的交互历史不是一种可扩展的解决方案:
> 标准的解决方法是将先前的上下文塞入上下文窗口。但这种方法在成本、延迟以及“中间遗忘”效应(有广泛的记录)等方面都存在缺陷,即模型会忽略远离提示词边缘的事实。100 万令牌的上下文窗口只是一个临时记事本,而非存储系统……真正缺失的是长期记忆:一种能够跨越会话终点、支持长达数年的交互,并允许按内容、时间和用户检索事实的持久化存储。
Atlas 的核心理念是认知科学界已经识别出的三种记忆类型:情景记忆,记录“发生了什么”;语义记忆,记录“什么是真实的”;程序性记忆,记录“什么有效”。由于每种记忆类型都有其自身的规则和生命周期,所以 Atlas 为每种记忆类型维护了单独的 Elasticsearch 索引。
记忆的生成是将用户的每个输入作为一个情景记忆事件进行存储。这些记忆大多会从记忆中消退,但其中一些会“成为持久的事实证据”。这是通过让大型语言模型(LLM)对其进行整合来实现的。LLM 将识别出新的事实或语义记忆,并将每项新事实以短句形式存储,同时附上作为证据的情景记忆,以及被该新事实所取代的任何前事实。
整合还会通过两种方式更新程序性记忆。首先,通过创建新的“操作指南”(即解决问题的一系列步骤);其次,更新现有操作指南的成功与失败计数。这些计数可能会影响检索结果,提升成功率更高的操作指南的优先级。
智能代理通过单一混合查询访问所有这些索引,并在 BM25 词汇搜索和 Jina v5 语义搜索的基础上应用了“互逆排名融合”(RRF)技术;合并后的结果会通过跨编码器重排序器进行重新排序。文档级安全(DLS)机制可以确保查询仅搜索属于该用户的记忆文档。
在 Hacker News 上 关于 Atlas 的一场讨论 中,有一些用户质疑,将 Elasticsearch 用作存储是否“大材小用”,并建议使用其他支持向量处理的数据库,例如 SQLite。另一位用户回复道:
> 一旦需要脚本评分之类的功能,“任何其他向量数据库”就会开始力不从心……这时问题就变成:“为了性能,你是否需要[近似最近邻]?”…… 诚然,对于数量远超大多数人预期的向量而言,暴力搜索的性能表现确实不错,但如果你希望其延迟达到网页级水平,那么在数量远低于 100 万时,它肯定会遇到瓶颈。维护 Elasticsearch 固然需要成本,但选择一个性能不足的数据库,随后又不得不将其移植到合适的数据库上,同样会耗费大量的时间。
Atlas 的源代码 已经在 GitHub 上开源。
原文链接: https://www.infoq.com/news/2026/06/elastic-atlas-agent-memory/
声明:本文由 InfoQ 翻译,未经许可禁止转载。
点击底部阅读原文访问 InfoQ 官网,获取更多精彩内容!****
今日好文推荐
Key Quotes
> The standard approach is to stuff prior context into the context window. However, this method suffers from drawbacks in cost, latency, and the 'middle forgetting' effect (widely documented), meaning the model ignores facts far from the prompt edges. A context window of 1 million tokens is merely a temporary notepad, not a storage system...... What is truly missing is long-term memory: a persistent store capable of spanning session endpoints, supporting interactions lasting years, and allowing retrieval of facts by content, time, and user.
> Once features like scripting scoring are needed, 'any other vector database' starts to fall short...... The question then becomes: 'For performance, do you need [approximate nearest neighbor]?'...... Indeed, for vector quantities far exceeding most people's expectations, brute-force search performance is actually quite good, but if you want latency at web-page level, it will surely hit a bottleneck well below a million vectors. Maintaining Elasticsearch certainly incurs cost, but choosing an inadequate database and later having to migrate to a suitable one also consumes significant time.
Tags
Elasticsearch
Intelligent Agent
Memory System
MCP
Hybrid Retrieval
Related Articles
* Tech Enthusiast Weekly (Issue 394): The Second Wave of API Openness * On Agent Memory * OpenClaw and Hermes: A Comprehensive Review of AI Agent Architecture from Source Code * JD.com Ad Large Model in Practice: How the GRAM Architecture Delivers Generative Recommendations in Under 50ms * Designing a Production-Grade Multi-Agent Harness from Scratch: Architecture, Evaluation, Memory, Cost, and MCP Tool Integration * CIOs Are Abandoning AI Code Generation Rate: A Practical Review of What Truly Counts as R&D Efficiency * From Requirements to Deployment: A Comprehensive Leap in R&D Efficiency with Full AI Automation to L3 (full automation), along with core challenges, specific practices, and quantified benefits.") * From Context to Experiential Assets: Engineering Path and Practice of Agent Memory Systems with MemOS * Integrating UI Generation into the Pipeline: Semi-Supervised Evaluation for Automated UI Production * After Consuming Hundreds of Billions of Tokens: Reflections on Software Construction, Forms, and Future Development in the Agent Era