GEO

AI如何实现永久记忆?Supermemory ASMR系统99%准确率解析

2026/3/25
AI如何实现永久记忆?Supermemory ASMR系统99%准确率解析
AI Summary (BLUF)

Supermemory's ASMR system achieves 99% accuracy on the LongMemEval benchmark by replacing traditional vector databases with a multi-agent parallel reasoning pipeline, enabling AI to maintain persistent memory across conversations.

原文翻译: Supermemory的ASMR系统通过用多智能体并行推理管道取代传统向量数据库,在LongMemEval基准测试中达到99%准确率,使AI能够在对话间保持持久记忆。

引言:AI 记忆难题迎来突破性进展

AI finally has "permanent memory"! Today, the super memory system ASMR made a grand debut, achieving a 99% score and shattering the SOTA record on the industry's most challenging AI memory benchmark. The entire internet is calling it insane.

AI 终于拥有了“永久记忆”!今天,超级记忆系统 ASMR 重磅登场,在业界公认最难的 AI 记忆基准测试中,取得了 99% 的准确率,刷新了 SOTA 记录。全网直呼太疯狂了。

AI 的记忆能力,尤其是长期、连贯且能处理时序与矛盾信息的记忆,一直是其迈向更高级智能体的核心障碍。传统的检索增强生成(RAG)方法在处理复杂、动态的对话历史时常常力不从心。然而,这一局面可能已被彻底改变。

The memory capability of AI, especially long-term, coherent memory that can handle temporal and contradictory information, has been a core obstacle in its evolution towards more advanced agents. Traditional Retrieval-Augmented Generation (RAG) methods often struggle with complex, dynamic conversation histories. However, this situation may have fundamentally changed.

核心突破:ASMR 系统与 LongMemEval 基准测试

什么是 ASMR?

ASMR (Agent Search and Memory Retrieval) is a novel memory system developed by the Supermemory team. It abandons traditional vector databases and embedding-based retrieval, operating entirely in memory through a multi-agent parallel inference pipeline.

ASMR(智能体搜索与记忆检索)是由 Supermemory 团队开发的一种新型记忆系统。它摒弃了传统的向量数据库和基于嵌入的检索方式,完全在内存中运行,通过多智能体并行推理流水线实现。

攻克最严苛的测试:LongMemEval

LongMemEval is one of the most rigorous publicly available benchmarks for long-term memory. It simulates real-world production scenarios with over 115,000 tokens of dialogue history, containing contradictory information, scattered events across multiple sessions, and complex questions requiring temporal reasoning.

LongMemEval 是目前公开可见的最严苛的长期记忆基准测试之一。它模拟了真实的生产环境,包含超过 11.5 万个 Token 的对话历史、相互矛盾的信息、跨越多个会话的零散事件,以及需要进行时间推理的复杂问题。

ASMR 正是在这一极具挑战性的测试中取得了接近完美的 99% 准确率,标志着 AI 在“健忘症”问题上取得了里程碑式的进展。

It was on this highly challenging benchmark that ASMR achieved a near-perfect 99% accuracy, marking a milestone in AI's battle against "amnesia."

技术架构:多智能体协同的革新设计

ASMR 的核心创新在于其“多智能体协同编排”架构,它用主动的智能体推理完全取代了传统的向量相似度计算。

The core innovation of ASMR lies in its "multi-agent collaborative orchestration" architecture, which completely replaces traditional vector similarity calculations with proactive agent reasoning.

第一阶段:并行信息摄取

First, a team of 3 parallel Observer Agents, powered by models like Gemini 2.0 Flash, concurrently reads raw conversation data. Their goal is to extract structured knowledge across six dimensions:

  • Personal Information / 个人信息
  • Preferences / 偏好
  • Events / 事件
  • Temporal Data / 时序数据
  • Information Updates / 信息更新
  • Assistant Information / 助手信息

首先,由 3 个并行的观察者智能体(由 Gemini 2.0 Flash 等模型驱动)组成的团队并发读取原始对话数据。它们的目标是从六个维度提取结构化知识:

  • 个人信息
  • 偏好
  • 事件
  • 时序数据
  • 信息更新
  • 助手信息

这些结构化的发现被原生存储并映射回源会话,为高效检索奠定基础。

These structured findings are natively stored and mapped back to their source sessions, laying the foundation for efficient retrieval.

第二阶段:主动推理检索

When a query is received, instead of querying a vector database, ASMR deploys 3 parallel Searcher Agents. Each agent has a specialized focus:

  • Agent 1: Searches for direct facts and explicit statements. / 搜索直接事实和明确陈述。
  • Agent 2: Looks for relevant context, social cues, and implications. / 寻找相关的上下文、社交线索和暗示。
  • Agent 3: Reconstructs timelines and relationship graphs. / 重建时间线和关系图谱。

当接收到查询时,ASMR 不会查询向量数据库,而是部署 3 个并行的搜索智能体。每个智能体都有专属的侧重点:

  • 智能体 1:搜索直接事实和明确陈述。
  • 智能体 2:寻找相关的上下文、社交线索和暗示。
  • 智能体 3:重建时间线和关系图谱。

编排器汇总所有搜索结果,并提取原文片段进行验证,实现了基于认知理解的智能检索。

The orchestrator aggregates all search results and extracts verbatim text fragments for verification, enabling intelligent retrieval based on cognitive understanding.

第三阶段:专业化答案生成

为了处理 LongMemEval 中多样化的问题,团队测试了两种高效的答案生成工作流:

To handle the diverse questions in LongMemEval, the team tested two efficient answer generation workflows:

1. 8-Variant Cluster (98.6% Accuracy): The retrieved context is routed to 8 highly specialized prompt variants (e.g., precise counter, time expert) running in parallel. If any variant produces the correct answer, the question is marked correct.

1. 8 变体集群(98.6% 准确率):检索到的上下文被路由给 8 个高度专业化的提示词变体(例如,精准计数器、时间专家)并行运行。如果任何一个变体得出正确答案,该问题即被标记为正确。

2. 12-Variant Decision Forest (97.2% Accuracy): 12 specialized AI agents (powered by GPT-4o-mini) answer independently. A final aggregator LLM acts as a judge, synthesizing answers through majority voting and conflict resolution to produce a single consensus answer.

2. 12 变体决策森林(97.2% 准确率):12 个专业化的 AI 智能体(由 GPT-4o-mini 驱动)独立回答问题。一个最终的聚合大模型充当裁判,通过多数投票和冲突解决机制综合答案,产生单一的共识答案。

超越实验:Supermemory 的宏大愿景

ASMR 不仅仅是一个刷榜的实验系统,它背后是 Supermemory 团队打造的一个完整的记忆与上下文基础设施。

ASMR is not just an experimental system for topping leaderboards. Behind it lies a complete memory and context infrastructure built by the Supermemory team.

记忆 vs. RAG:本质区别

Supermemory aims to endow AI with true memory, not just retrieval. The key difference is that RAG is user-agnostic and static, while Supermemory actively extracts facts, tracks changes, resolves contradictions, and even implements automatic forgetting of temporary information.

Supermemory 的目标是赋予 AI 真正的记忆,而不仅仅是检索能力。关键区别在于,RAG 不区分用户且是静态的,而 Supermemory 会主动提取事实、追踪变化、处理矛盾,甚至实现临时信息的自动遗忘。

核心产品能力

  • Unified Memory & RAG API: Combines knowledge base retrieval and personalized context in a single API call (~50ms latency).

    统一的记忆与 RAG API:在一次 API 调用中(约 50 毫秒延迟)合并知识库检索和个性化上下文。

  • Automated User Profiling: Automatically builds and maintains user profiles (static facts & dynamic context), enabling AI agents to switch to "old friend mode."

    自动化用户画像:自动构建和维护用户画像(静态事实和动态上下文),使 AI 智能体能够切换到“老朋友模式”。

  • Extensive Connectivity: Supports automatic syncing via webhooks with external data sources like Google Drive, Notion, GitHub, etc., with built-in processing for various file types.

    广泛的连接能力:支持通过 Webhook 与 Google Drive、Notion、GitHub 等外部数据源自动同步,并内置了对多种文件类型的处理能力。

  • Developer-Friendly Integration: Offers SDKs for popular frameworks (Vercel AI SDK, LangChain, etc.) and an MCP server for easy integration into IDEs like Cursor and VS Code.

    开发者友好集成:为流行框架(Vercel AI SDK、LangChain 等)提供 SDK,并提供 MCP 服务器以便轻松集成到 Cursor 和 VS Code 等 IDE 中。

结论:记忆——AI 进化的下一块拼图

The competition in AI has long focused on model scale, inference speed, and context length. However, a large context window alone does not equate to memory. True memory is the final piece that transforms AI from a "tool" into a "partner." When every agent can remember who you are and what you were doing, human-computer interaction will undergo a quiet yet profound transformation.

AI 的竞争长期以来集中在模型规模、推理速度和上下文长度上。然而,仅有一个大的上下文窗口并不等同于记忆。真正的记忆是将 AI 从“工具”转变为“伙伴”的最后一块拼图。当每个智能体都能记住你是谁、你之前在做什么时,人机交互将经历一场静默而深刻的变革。

Supermemory 团队通过 ASMR 系统及其产品化实践,正致力于将“工作记忆”从 AI 的附加功能提升为一层核心基础设施。随着其宣布将于 4 月初开源 ASMR 全部代码,AI 记忆的“大航海时代”或许真的即将开启。

By developing the ASMR system and its productized practices, the Supermemory team is working to elevate "working memory" from an add-on feature of AI to a core infrastructure layer. With the announcement that ASMR's full code will be open-sourced in early April, the "Age of Exploration" for AI memory may truly be about to begin.


本文基于公开报道及 Supermemory 团队发布的信息进行分析。所有图片来源均为原报道。

常见问题(FAQ)

Supermemory的ASMR系统与传统RAG方法有什么根本区别?

ASMR系统摒弃了传统的向量数据库和基于嵌入的检索,采用多智能体并行推理管道在内存中运行,通过主动的智能体推理取代向量相似度计算,从而实现了基于认知理解的智能检索。

ASMR系统在LongMemEval基准测试中表现如何?为什么这个测试很重要?

ASMR在LongMemEval基准测试中取得了99%的准确率。该测试模拟真实生产环境,包含超过11.5万Token的对话历史、矛盾信息和需要时间推理的复杂问题,是业界公认最严苛的AI长期记忆测试。

ASMR系统的多智能体架构是如何工作的?

系统分为三个阶段:1)并行信息摄取:3个观察者智能体从六个维度提取结构化知识;2)主动推理检索:3个搜索智能体分别专注于事实、上下文和时间线;3)专业化答案生成:通过智能体协作生成准确回答。

← 返回文章列表
分享到:微博

版权与免责声明:本文仅用于信息分享与交流,不构成任何形式的法律、投资、医疗或其他专业建议,也不构成对任何结果的承诺或保证。

文中提及的商标、品牌、Logo、产品名称及相关图片/素材,其权利归各自合法权利人所有。本站内容可能基于公开资料整理,亦可能使用 AI 辅助生成或润色;我们尽力确保准确与合规,但不保证完整性、时效性与适用性,请读者自行甄别并以官方信息为准。

若本文内容或素材涉嫌侵权、隐私不当或存在错误,请相关权利人/当事人联系本站,我们将及时核实并采取删除、修正或下架等处理措施。 也请勿在评论或联系信息中提交身份证号、手机号、住址等个人敏感信息。