加利福尼亚大学圣迭戈分校博士生王禹和纽约大学教授陈溪联合推出并开源了 MIRIX，全球首个真正意义上的多模态、多智能体AI记忆系 统。MIRIX团队同步上线了一款桌面端APP，可直接下载使用！ 还记得第一次用 GPT 写邮件的惊喜吗？却也一定遇到过今天的 AI「忘性」——聊得再深入，窗口一关，历史烟消云散。 因此，研究人员认为：从「对话」到「记忆」，将是AI进化的必经之路。 研究人员推出并开源MIRIX，全球首个真正意义上的多模态、多智能体AI记忆系统。 在ScreenshotVQA这一需要深度多模态理解的挑战性基准上，MIRIX的准确率比传统RAG方法高出35%，存储开销降低99.9%，与长文本方法相比超出 410%，开销降低93.3%。 在LOCOMO长对话任务中，MIRIX以85.4%的成绩显著超越所有现有方法，树立了新的性能标杆。 与此同时，研究人员在Mac端上线了一款应用产品，通过这款开箱即用的应用程序，终于可以为每个人构建专属于自己的AI个人助理。 桌面端APP使用场景 直接访问官方网站，即可直接下载APP: 论文链接：https://arxiv.org/abs/2507.07957 官方网站：h ...

Core Insights - The article discusses the evolution of AI from being a mere text generator to an intelligent agent with memory capabilities, emphasizing the need for a systematic understanding of AI memory mechanisms in the context of large models [1][2][4] Summary by Sections AI Memory Framework - A systematic framework for AI memory is constructed based on two dimensions: operation and representation [4] - Memory representation is categorized into parametric memory and contextual memory, with six fundamental memory operations identified: consolidation, updating, indexing, forgetting, retrieval, and compression [5][6] Memory Operations - Memory management operations control the storage, maintenance, and pruning of information, ensuring the evolution of system memory over time [12] - Key operations include: - Consolidation: Transforming short-term experiences into long-term memory [26] - Indexing: Creating structured access paths to enhance retrieval efficiency [12] - Updating: Modifying existing memory based on new knowledge [13] - Forgetting: Selectively removing outdated or harmful memory content [14] Memory Utilization - Memory utilization refers to how models access and use stored information during inference, including retrieval and compression operations [15] - Retrieval involves identifying relevant memory segments based on input [15] - Compression retains key information while discarding redundant content, crucial for efficient memory utilization [16] Key Research Themes - The article identifies four core themes in AI memory research: - Long-term memory: Focuses on cross-session memory management and personalized reasoning [19] - Long-context memory: Addresses efficiency in handling extensive contextual information [19] - Parametric memory modification: Involves dynamic rewriting of internal knowledge [19] - Multi-source memory integration: Emphasizes the unification of diverse data sources for robust semantic understanding [19] Practical Applications - AI memory integration is becoming essential for various applications, including programming assistants, personalized recommendations, and structured intelligent agents [50] - Notable products like ChatGPT and GitHub Copilot illustrate the shift from task-oriented tools to long-term partners in user interaction [50] Future Directions - The article highlights the need for breakthroughs in memory mechanisms to achieve long-term adaptation, cross-modal understanding, and personalized reasoning in AI systems [55] - Key challenges include unified evaluation of long-term memory, efficient long-context modeling, and conflict detection in multi-source memory systems [55]