多智能体如何超越向量检索？2026年AI记忆理解新突破

How Does Agent Collaboration Achieve Memory Understanding Beyond Vector Retrieval?

快速阅读 / TL;DR

Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。团队利用多智能体协作系统，在长期记忆基准测试LongMemEval长期记忆基准测试，包含11.5万token的对话历史，用于评估AI系统的记忆能力。上达到了99%的准确率。其核心突破在于使用3个并行的搜索Agent搜索代理，在检索阶段专门负责不同任务的Agent：事实查找、语境挖掘、时间线重建，通过主动阅读和推理而非向量计算来检索信息。替代了传统的向量检索，使AI能够通过“理解”而非“数学相似度”来回忆信息。这套方案无需向量数据库，甚至可以嵌入到机器人等设备中运行。

The Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。 team achieved 99% accuracy on the LongMemEval长期记忆基准测试，包含11.5万token的对话历史，用于评估AI系统的记忆能力。 long-term memory benchmark using a multi-agent collaboration system. The core breakthrough lies in replacing traditional vector retrieval with three parallel search Agents, enabling AI to recall information through "understanding" rather than "mathematical similarity." This solution does not require a vector database and can even be embedded into devices like robots.

引言：向量检索的局限性 / Introduction: The Limitations of Vector Retrieval

向量数据库可能不是AI记忆的最优解。传统检索增强生成（RAG）的问题往往出在检索环节。基于语义相似度的匹配难以区分“旧事实”和“新更正”，当检索结果中混杂过多噪音时，大型语言模型就容易迷失方向，导致回答不准确或自相矛盾。

Vector databases may not be the optimal solution for AI memory. The problem with traditional Retrieval-Augmented Generation (RAG) often lies in the retrieval stage. Semantic similarity-based matching struggles to distinguish between "old facts" and "new corrections." When retrieval results are contaminated with excessive noise, large language models can become confused, leading to inaccurate or contradictory responses.

核心突破：ASMR架构 / The Core Breakthrough: The ASMR Architecture

Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。团队的解法是ASMR（Agentic Search and Memory Retrieval，智能体搜索与记忆检索）。这套架构完全抛弃了向量检索，转而采用多智能体协作来模拟人类的“理解式”回忆过程。

The solution from the Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。 team is ASMR (Agentic Search and Memory Retrieval). This architecture completely abandons vector retrieval, instead employing multi-agent collaboration to simulate a human-like "understanding-based" recall process.

信息摄取阶段 / Information Ingestion Phase

在信息摄取阶段，3个并行的Observer Agent观察者代理，在信息摄取阶段并行读取对话记录，按照个人信息、偏好、事件、时间数据等六个维度提取知识点。同时读取对话记录。它们按照六个维度（如个人信息、偏好、事件、时间数据等）提取知识点，并直接存储结构化的内容，而非生成向量嵌入（embedding）。

During the information ingestion phase, three parallel Observer Agent观察者代理，在信息摄取阶段并行读取对话记录，按照个人信息、偏好、事件、时间数据等六个维度提取知识点。s simultaneously read conversation records. They extract knowledge points according to six dimensions (such as personal information, preferences, events, temporal data, etc.) and directly store structured content, rather than generating vector embeddings.

检索阶段：主动推理取代被动匹配 / Retrieval Phase: Active Reasoning Replaces Passive Matching

检索阶段是ASMR的关键。面对用户提问时，系统不再查询向量数据库，而是派出3个专门的搜索Agent搜索代理，在检索阶段专门负责不同任务的Agent：事实查找、语境挖掘、时间线重建，通过主动阅读和推理而非向量计算来检索信息。进行协同工作：

• 事实搜索Agent搜索代理，在检索阶段专门负责不同任务的Agent：事实查找、语境挖掘、时间线重建，通过主动阅读和推理而非向量计算来检索信息。：负责查找直接相关的事实。
• 语境挖掘Agent：负责挖掘对话中隐含的上下文和关联信息。
• 时间线重建Agent：负责梳理事件的时间顺序和逻辑关系。

这些Agent本质上是在进行“主动阅读和推理”，而不是执行被动的向量余弦相似度计算。

The retrieval phase is the key to ASMR. When faced with a user query, the system no longer queries a vector database. Instead, it dispatches three specialized search Agents for collaborative work:

• Fact Search Agent: Responsible for finding directly related facts.
• Context Mining Agent: Responsible for挖掘 implicit context and关联 information within the conversation.
• Timeline Reconstruction Agent: Responsible for梳理 the chronological order and logical relationships of events.

These Agents are essentially engaged in "active reading and reasoning," rather than performing passive vector cosine similarity calculations.

回答生成策略 / Answer Generation Strategies

团队测试了两种创新的回答生成策略：

1. 并行专家Prompt策略：使用8个高度专业化的提示词变体（如精确计数专家、时间专家、上下文深挖专家等）并行运行。只要其中任何一条推理路径得出正确答案，即视为成功。该策略准确率达到98.6%。
2. 多Agent投票裁决策略：12个独立的Agent分别生成答案，随后由一个聚合器LLM聚合器大语言模型，在回答阶段综合多个Agent的投票结果做出最终裁决。综合所有答案进行投票裁决。该策略准确率达到97.2%。

The team tested two innovative answer generation strategies:

1. Parallel Expert Prompt Strategy: Eight highly specialized prompt variants (e.g., precise counting expert, time expert, context deep-dive expert) run in parallel. Success is achieved if any of these reasoning paths produces the correct answer. This strategy achieved an accuracy of 98.6%.
2. Multi-Agent Voting Adjudication Strategy: Twelve independent Agents generate answers separately, followed by an aggregator LLM synthesizing all answers for a final vote. This strategy achieved an accuracy of 97.2%.

主要分析与启示 / Main Analysis and Implications

从“数学相似”到“认知理解” / From "Mathematical Similarity" to "Cognitive Understanding"

这一成果的核心启示在于，它证明了在处理复杂的记忆任务时，“认知理解”可能比“数学相似性”更为有效。数学相似性主要捕捉文本的表层模式和统计特征，而智能体协作系统能够主动处理时间序列中的矛盾、信息更新以及语义上的细微差别，更接近人类的理解过程。

The core insight from this achievement is that it demonstrates "cognitive understanding" may be more effective than "mathematical similarity" for handling complex memory tasks. Mathematical similarity primarily captures surface-level patterns and statistical features of text, whereas an agent collaboration system can actively handle contradictions in timelines, information updates, and semantic nuances—closer to the human process of understanding.

轻量化与普适性 / Lightweight and Universal Architecture

另一个显著优势是该架构的轻量化与普适性。ASMR系统完全在内存中运行，不依赖任何外部向量数据库。这不仅降低了系统复杂性和延迟，也意味着理论上它可以被部署到任何计算设备上，包括资源受限的嵌入式系统或机器人，为AI的泛在化部署提供了新的可能性。

Another significant advantage is the architecture's lightweight and universal nature. The ASMR system runs entirely in memory, without relying on any external vector database. This not only reduces system complexity and latency but also意味着 it can theoretically be deployed on any computing device, including resource-constrained embedded systems or robots, offering new possibilities for the ubiquitous deployment of AI.

结论与展望 / Conclusion and Outlook

当未来数十亿个高度个性化的AI Agent开始持续学习和记忆与我们相关的一切时，记忆系统的性能天花板在哪里？Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。团队的工作提示我们，答案或许不在于无限堆叠的算力，而在于我们愿意赋予这些智能体多少“主动思考”与“协作理解”的权限。通过模拟更高级的认知过程，我们有望构建出更可靠、更人性化的AI记忆系统。

As billions of highly personalized AI Agents in the future begin to continuously learn and remember everything about us, where is the performance ceiling for memory systems? The work of the Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。 team suggests that the answer may not lie in endlessly stacking computational power, but rather in how much permission we are willing to grant these agents for "active thinking" and "collaborative understanding." By simulating higher-level cognitive processes, we have the potential to build more reliable and human-like AI memory systems.

参考 / Reference: x.com/DhravyaShah/status/2035517012647272689

常见问题（FAQ）

Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。的ASMR架构如何解决传统向量检索的噪音问题？

ASMR架构用3个并行搜索Agent搜索代理，在检索阶段专门负责不同任务的Agent：事实查找、语境挖掘、时间线重建，通过主动阅读和推理而非向量计算来检索信息。（事实搜索、语境挖掘、时间线重建）替代向量检索，通过主动推理理解信息关联，避免语义相似度匹配带来的噪音干扰。

为什么Supermemory一种基于多智能体协作的记忆系统，通过Agent的认知理解替代传统向量检索，在LongMemEval基准测试中达到99%准确率。方案不需要向量数据库？

该系统在信息摄取阶段直接存储Observer Agent观察者代理，在信息摄取阶段并行读取对话记录，按照个人信息、偏好、事件、时间数据等六个维度提取知识点。提取的结构化内容（六个维度知识点），检索时通过多Agent协作进行理解式回忆，完全绕过了向量嵌入和相似度计算。

多智能体协作如何实现99%的准确率？

采用并行专家Prompt策略（8个专业提示词并行运行）或多Agent投票裁决策略，让多个推理路径协同工作，只要任一路径得出正确答案即视为成功，极大提升了召回精度。