第八章 · Section 8
Long-Term Memory Storage (MS) · 长时记忆存储
Long-Term Memory Storage (MS): The ability to stably acquire, consolidate, and store new information from recent experiences.
长时记忆存储(MS):稳定地获取、巩固并存储来自近期经验的新信息的能力。
The ability to link previously unrelated pieces of information. Remember connections between text, images, audio.
• Cross-Modal Association — form a link between two previously unrelated stimuli.
Sample: "You met this person yesterday, what was her name?"
• Personalization Adherence — remember and apply user preferences.
Sample: "Sign off my emails as I usually do." / "Please recall the address I mentioned earlier today."
• Procedural Association — remember and execute a sequence of steps or rules.
Sample: "Please format the Balance Sheet to match the new standard discussed this week."
把先前无关的信息片段联系起来的能力。记住文本、图像、音频之间的联系。
• 跨模态联想——在先前无关的两个刺激间建立联结。
样例:"你昨天见过这个人,她叫什么名字?"
• 个性化遵从——记住并应用用户偏好。
样例:"像往常一样给我的邮件署名。" / "请回忆起我今天早些时候提到的地址。"
• 程序性联想——记住并执行一串步骤或规则。
样例:"请把资产负债表按本周讨论的新标准格式化。"
The ability to encode and recall the semantic gist of experiences and narratives.
• Story Recall — remember the gist of stories.
Sample: "Please summarize the ending of my novel draft from yesterday."
• Movie Recall — remember the gist of movies.
Sample: "What was the main conflict in the movie I showed to you last weekend?"
• Episodic Context Recall — remember specific events and experiences, including their context (the "what, where, when, and how").
Sample: "What topic did we discuss yesterday with her?"
编码并回忆经验与叙事的语义要点(gist)的能力。
• 故事回忆——记住故事的要点。
样例:"请概括我昨天的小说草稿的结尾。"
• 电影回忆——记住电影的要点。
样例:"我上周末给你看的电影,主要冲突是什么?"
• 情景语境回忆——记住特定事件与经验及其语境("什么、何地、何时、如何")。
样例:"我们昨天和她讨论了什么话题?"
The ability to store and reproduce information precisely as it was presented.
• Short Sequence Recall — exactly recall short sequences of text after a delay.
• Set Recall — recall a set (order of recall does not matter).
• Design Recall — remember a design pattern / the spatial arrangement and structure of visual information.
Sample: "Can you recreate the simple layout we reviewed yesterday?"
精确按原样存储并重现信息的能力。
• 短序列回忆——延迟后精确回忆短文本序列。
• 集合回忆——回忆一个集合(回忆顺序无关)。
• 图案回忆——记住一个设计图案 / 视觉信息的空间排布与结构。
样例:"你能重建我们昨天看过的那个简单布局吗?"
评估细节与 AI 表现
Assessment Details & AI Performance
Assessment Details. See Appendix F for further details on how to assess long-term memory storage capabilities concretely.
AI System Performance. The table summarizes current AI system performance on Long-Term Memory Storage (MS) tasks. Both GPT-4 and GPT-5 lack appreciable long-term memory storage capabilities.
Model | Associative (4%) | Meaningful (3%) | Verbatim (3%) | Total
GPT-4 | 0% | 0% | 0% | 0%
GPT-5 | 0% | 0% | 0% | 0%
评估细节:如何在具体层面评估长时记忆存储,参见附录 F。
AI 系统表现:下表汇总当前 AI 系统在长时记忆存储(MS)任务上的表现。GPT-4 与 GPT-5 都缺乏可观的长时记忆存储能力。
模型 | 联想(4%) | 意义(3%) | 逐字(3%) | 合计
GPT-4 | 0% | 0% | 0% | 0%
GPT-5 | 0% | 0% | 0% | 0%