Core Insights - Google has introduced a new framework called HOPE to address the long-term memory issues in large models, which has been a significant challenge affecting the depth and breadth of AI applications [1][2][4] Group 1: Long-term Memory Challenges - Long-term memory is crucial for AI to function as a "persistent assistant" rather than a one-time use tool, impacting its ability to remember key details across different tasks [2][4] - The Titans architecture proposed by Google last year has been a focal point in discussions about long-term memory, emphasizing the need for a sustainable memory component rather than merely extending context windows [4][9] Group 2: Recent Developments in AI Assistants - Google has launched significant updates for Gemini, including an "automatic memory" feature that learns from past conversations to provide personalized responses [5] - Other leading AI assistants, such as ChatGPT and iFlytek's Xunfei Spark, are also integrating long-term memory modules to maintain continuity across conversations and tasks [5][12] Group 3: Evolution of Memory Mechanisms - The understanding of long-term memory is shifting from merely storing text to retaining experiences that influence decision-making processes [11][19] - The introduction of frameworks like Evo-Memory benchmark and ReMem aims to integrate long-term memory into the workflow of intelligent agents, assessing their ability to extract and utilize experiences in continuous tasks [11][12] Group 4: Industry Comparisons - Different approaches to long-term memory are emerging within the industry, such as MiniMax's focus on linear attention architecture and DeepSeek's externalized memory components [16][19] - The emphasis is on creating a memory mechanism that is not just a passive storage solution but actively participates in decision-making, reflecting a significant shift in the role of long-term memory in AI models [20]

可以说，大模型的「短期能力」决定了它能不能把一句话说通，但长期记忆真正决定的，其实是它有没有资格被称为「助手」。 也正是因为这一点，去年最后一天谷歌研究团队提出的 Titans 架构，在 2025 年被反复翻出来讨论，并不意外。这篇论文试图回答的，并不是「上下文还能 拉多长」这种老问题，而是一个更本质的命题： 当注意力只是短期记忆，大模型到底该如何拥有真正的长期记忆。 日前，Google在其发布的论文《Nested Learning: The Illusion of Deep Learning Architectures》中，提出了一个名为 HOPE 的新框架试图解决大模型长期记忆 的问题。 这一架构备受关注，因为长期记忆一直困扰着大模型的发展，甚至影响着AI落地到智能体的广度与深度。 今天让 AI 写一段漂亮的回答不难，难的是隔了一周、换了工作任务，它还记得你之前某次对话的关键细节，不断更新对你的个性化记忆。也只有在这一 刻，大模型才真正开始接近「持续工作的智能体」，而不是一次性消耗品。 图片来源：谷歌 在 Titans 里，Transformer 的 self-attention（自注意力机制）被明确界定 ...