Do you need to file a feature request?
Feature Request Description
RFC：关于DeepTutor下一代Memory架构思考
1. 背景
DeepTutor 已经具备基础 memory 能力，例如用户画像、学习摘要、TutorBot 私有记忆、会话恢复、Notebook、Question Bank、Knowledge Hub 等。
但从使用者角度看，当前 memory 更像是在回答：
用户之前聊过什么？
用户大概喜欢什么学习风格？
用户当前有哪些学习资料？
而真正的个性化学习系统还需要回答：
用户真正掌握了什么？
哪些知识点只是熟悉但不会独立使用？
哪些错误反复出现？
哪些前置知识正在卡住后续学习？
下一步应该学什么、复习什么、练什么？
为什么系统这么判断？
因此，DeepTutor 的 memory 不应该只是聊天摘要，而应该升级为“学习者状态系统”。
2. 核心问题
当前 memory 主要缺少四类能力。
第一，缺少概念级学习状态。
系统需要知道用户在每个知识点上的掌握情况，而不是只知道“用户学过概率论”或者“用户最近问过贝叶斯公式”。
例如：
贝叶斯公式：
- 能识别公式
- 能复述概念
- 条件反转题容易错
- 迁移到医学检测场景时不稳定
- 需要 3 天内复习
第二，缺少错误模式记忆。
教学系统最有价值的 memory 不是“你错过哪道题”，而是“你为什么反复错”。
例如：
用户不是不会条件概率，而是经常混淆 P(A|B) 和 P(B|A)。
用户不是不会链式法则，而是在嵌套函数中经常漏掉内层导数。
第三，缺少跨模块的学习事件沉淀。
DeepTutor 有 Chat、Deep Solve、Quiz、Notebook、Book、Research、TutorBot 等模块。每个模块都会产生学习信号，但这些信号需要统一进入 memory。
例如：
Quiz 答错了某个知识点
Deep Solve 中请求了多次提示
Notebook 保存了一段自己的理解
Book 中完成了一页学习
TutorBot 发起了复习提醒
这些都应该被视为学习证据，而不是散落在不同功能里的日志。
第四，缺少可解释和可修正的 memory。
如果系统判断“用户不懂某个概念”，用户应该能看到依据，并可以修改或删除错误判断。
例如：
系统判断：你在条件概率上存在混淆。
依据：
1. 上次答题时反转了 P(A|B) 和 P(B|A)
2. 解释时没有区分条件和结果
3. 提示后能够修正
用户可以选择：
确认 / 修改 / 删除 / 重新测试
3. 目标
本 RFC 的目标是把 DeepTutor memory 从“记住用户信息”升级为“支撑个性化教学决策”。
具体目标包括：
1. 统一所有学习行为的记录方式
2. 建立概念级学习状态
3. 记录长期错误模式
4. 支持复习和遗忘管理
5. 让 TutorBot 能根据用户状态调整教学方式
6. 让 Notebook、Quiz、Book、Deep Solve 都能反哺 memory
7. 让用户能查看、纠正、删除系统记忆
8. 为后续学习插件提供统一 memory 基础
4. 建议的 Memory 分层
DeepTutor 的 memory 可以分成六层。
4.1 Profile Memory：用户画像
记录用户相对稳定的信息。
包括：学习目标，知识背景，偏好的解释风格，语言偏好，学习节奏，当前关注方向
这部分 DeepTutor 已经有基础能力，可以继续保留。
4.2 Summary Memory：学习摘要
记录用户近期和长期的学习轨迹。
包括：最近学了什么，完成了哪些任务，探索过哪些主题，阶段性理解如何变化
4.3 Learner State Memory：学习状态
这是新增的核心层。
它记录用户在每个知识点上的状态。
例如：
概念：贝叶斯公式
识别能力：较好
回忆能力：一般
应用能力：一般
迁移能力：较弱
解释能力：较弱
最近错误：混淆 P(A|B) 和 P(B|A)
最近复习：3 天前
下一次复习：明天
这层 memory 决定 DeepTutor 是否真正能做到个性化教学。
4.4 Misconception Memory：误解与错误模式
记录用户反复出现的错误。
例如：
错误模式：条件概率方向混淆
出现次数：4 次
最近出现：昨天
关联概念：条件概率、贝叶斯公式、独立性
修复建议：先做条件方向判断题，再做贝叶斯应用题
这层 memory 可以让系统避免反复泛泛解释，而是直接修正关键误解。
4.5 Artifact Memory：学习资产记忆
把 Notebook、Question Bank、Book、Research notes 等学习产物纳入 memory。
它要解决的问题是：
用户保存的笔记能不能影响后续讲解？
用户做错的题能不能进入复习计划？
用户读过的章节能不能影响下一次 Quiz？
用户写过的解释能不能用来判断理解程度？
Artifact 不只是文件，而是学习证据。
4.6 TutorBot Private Memory：TutorBot 私有记忆
每个 TutorBot 可以有自己的记忆。
例如：
数学 TutorBot 记住用户最近在微积分上卡住。
写作 TutorBot 记住用户偏好先给结构再润色。
研究 TutorBot 记住用户正在跟踪某个论文方向。
RFC: Thoughts on the Next-Generation Memory Architecture for DeepTutor
1. Background
DeepTutor already has basic memory capabilities, such as user profiles, learning summaries, TutorBot private memory, session recovery, Notebook, Question Bank, Knowledge Hub, and so on.
However, from the user’s perspective, the current memory is mainly answering questions such as:
What has the user discussed before?
What kind of learning style does the user roughly prefer?
What learning materials does the user currently have?
A truly personalized learning system also needs to answer:
What has the user actually mastered?
Which concepts are merely familiar but cannot yet be used independently?
Which mistakes appear repeatedly?
Which prerequisite knowledge is blocking further learning?
What should the user learn, review, or practice next?
Why does the system make this judgment?
Therefore, DeepTutor’s memory should not merely be a chat summary. It should be upgraded into a learner state system.
2. Core Problems
The current memory mainly lacks four types of capabilities.
First, it lacks concept-level learning state. The system needs to know the user’s mastery status for each knowledge point, instead of only knowing that “the user has studied probability theory” or “the user recently asked about Bayes’ theorem.”
For example:
Bayes’ theorem:
Can recognize the formula.
Can restate the concept.
Often makes mistakes on condition-reversal questions.
Is unstable when transferring the concept to medical testing scenarios.
Needs review within three days.
Second, it lacks memory of error patterns. The most valuable memory for a tutoring system is not “which question the user got wrong,” but “why the user keeps making the same mistake.”
For example:
The user is not unable to understand conditional probability, but often confuses P(A|B) with P(B|A).
The user is not unable to understand the chain rule, but often misses the inner derivative in nested functions.
Third, it lacks cross-module consolidation of learning events. DeepTutor has modules such as Chat, Deep Solve, Quiz, Notebook, Book, Research, and TutorBot. Each module generates learning signals, but these signals need to enter memory in a unified way.
For example:
The user answered a quiz question incorrectly on a specific knowledge point.
The user requested multiple hints in Deep Solve.
The user saved a piece of personal understanding in Notebook.
The user completed a page in Book.
TutorBot initiated a review reminder.
All of these should be treated as learning evidence, rather than logs scattered across different features.
Fourth, it lacks explainable and correctable memory. If the system judges that “the user does not understand a concept,” the user should be able to see the evidence and modify or delete incorrect judgments.
For example:
System judgment: You have confusion around conditional probability.
Evidence:
In the last quiz, you reversed P(A|B) and P(B|A).
In your explanation, you did not distinguish between the condition and the outcome.
You were able to correct the answer after receiving a hint.
The user can choose:
Confirm, modify, delete, or retest.
3. Goals
The goal of this RFC is to upgrade DeepTutor memory from “remembering user information” to “supporting personalized teaching decisions.”
Specific goals include:
Unify the recording format for all learning behaviors.
Build concept-level learning states.
Record long-term error patterns.
Support review and forgetting management.
Enable TutorBot to adjust its teaching style based on user state.
Allow Notebook, Quiz, Book, and Deep Solve to feed back into memory.
Allow users to view, correct, and delete system memories.
Provide a unified memory foundation for future learning plugins.
Suggested Memory Layers
4. Recommend Memory Archteciture for DeepTutor
DeepTutor’s memory can be divided into six layers.
4.1 Profile Memory: User Profile
This layer records relatively stable information about the user.
It includes learning goals, knowledge background, preferred explanation style, language preference, learning pace, and current areas of focus.
DeepTutor already has basic capabilities in this area, which can be retained.
4.2 Summary Memory: Learning Summary
This layer records the user’s recent and long-term learning trajectory.
It includes what the user has recently learned, which tasks have been completed, which topics have been explored, and how the user’s understanding has changed over time.
4.3 Learner State Memory: Learning State
This is the new core layer.
It records the user’s state for each knowledge point.
For example:
Concept: Bayes’ theorem
Recognition ability: good
Recall ability: average
Application ability: average
Transfer ability: weak
Explanation ability: weak
Recent mistake: confused P(A|B) with P(B|A)
Last review: three days ago
Next review: tomorrow
This memory layer determines whether DeepTutor can truly provide personalized teaching.
4.4 Misconception Memory: Misconceptions and Error Patterns
This layer records mistakes that repeatedly appear for the user.
For example:
Error pattern: confusion about the direction of conditional probability
Occurrences: four times
Most recent occurrence: yesterday
Related concepts: conditional probability, Bayes’ theorem, independence
Repair suggestion: first practice identifying the direction of conditions, then practice Bayes application problems
This memory layer allows the system to avoid giving generic explanations repeatedly and instead directly correct the key misconception.
4.5 Artifact Memory: Learning Asset Memory
This layer incorporates learning artifacts such as Notebook, Question Bank, Book, and Research notes into memory.
It aims to answer the following questions:
Can the notes saved by the user influence future explanations?
Can the questions the user got wrong enter the review plan?
Can the chapters the user has read influence the next Quiz?
Can the user’s own explanations be used to judge their level of understanding?
Artifacts are not merely files. They are learning evidence.
4.6 TutorBot Private Memory: TutorBot Private Memory
Each TutorBot can have its own memory.
For example:
A Math TutorBot remembers that the user has recently been stuck on calculus.
A Writing TutorBot remembers that the user prefers getting the structure first, then polishing.
A Research TutorBot remembers that the user is tracking a particular research direction.
Related Module
Other
Use Case
No response
Additional Context
No response
Do you need to file a feature request?
Feature Request Description
RFC：关于DeepTutor下一代Memory架构思考
1. 背景
DeepTutor 已经具备基础 memory 能力，例如用户画像、学习摘要、TutorBot 私有记忆、会话恢复、Notebook、Question Bank、Knowledge Hub 等。
但从使用者角度看，当前 memory 更像是在回答：
而真正的个性化学习系统还需要回答：
因此，DeepTutor 的 memory 不应该只是聊天摘要，而应该升级为“学习者状态系统”。
2. 核心问题
当前 memory 主要缺少四类能力。
第一，缺少概念级学习状态。
系统需要知道用户在每个知识点上的掌握情况，而不是只知道“用户学过概率论”或者“用户最近问过贝叶斯公式”。
例如：
第二，缺少错误模式记忆。
教学系统最有价值的 memory 不是“你错过哪道题”，而是“你为什么反复错”。
例如：
第三，缺少跨模块的学习事件沉淀。
DeepTutor 有 Chat、Deep Solve、Quiz、Notebook、Book、Research、TutorBot 等模块。每个模块都会产生学习信号，但这些信号需要统一进入 memory。
例如：
这些都应该被视为学习证据，而不是散落在不同功能里的日志。
第四，缺少可解释和可修正的 memory。
如果系统判断“用户不懂某个概念”，用户应该能看到依据，并可以修改或删除错误判断。
例如：
3. 目标
本 RFC 的目标是把 DeepTutor memory 从“记住用户信息”升级为“支撑个性化教学决策”。
具体目标包括：
4. 建议的 Memory 分层
DeepTutor 的 memory 可以分成六层。
4.1 Profile Memory：用户画像
记录用户相对稳定的信息。
包括：学习目标，知识背景，偏好的解释风格，语言偏好，学习节奏，当前关注方向
这部分 DeepTutor 已经有基础能力，可以继续保留。
4.2 Summary Memory：学习摘要
记录用户近期和长期的学习轨迹。
包括：最近学了什么，完成了哪些任务，探索过哪些主题，阶段性理解如何变化
4.3 Learner State Memory：学习状态
这是新增的核心层。
它记录用户在每个知识点上的状态。
例如：
这层 memory 决定 DeepTutor 是否真正能做到个性化教学。
4.4 Misconception Memory：误解与错误模式
记录用户反复出现的错误。
例如：
这层 memory 可以让系统避免反复泛泛解释，而是直接修正关键误解。
4.5 Artifact Memory：学习资产记忆
把 Notebook、Question Bank、Book、Research notes 等学习产物纳入 memory。
它要解决的问题是：
Artifact 不只是文件，而是学习证据。
4.6 TutorBot Private Memory：TutorBot 私有记忆
每个 TutorBot 可以有自己的记忆。
RFC: Thoughts on the Next-Generation Memory Architecture for DeepTutor
1. Background
DeepTutor already has basic memory capabilities, such as user profiles, learning summaries, TutorBot private memory, session recovery, Notebook, Question Bank, Knowledge Hub, and so on.
However, from the user’s perspective, the current memory is mainly answering questions such as:
A truly personalized learning system also needs to answer:
Therefore, DeepTutor’s memory should not merely be a chat summary. It should be upgraded into a learner state system.
2. Core Problems
The current memory mainly lacks four types of capabilities.
First, it lacks concept-level learning state. The system needs to know the user’s mastery status for each knowledge point, instead of only knowing that “the user has studied probability theory” or “the user recently asked about Bayes’ theorem.”
For example:
Second, it lacks memory of error patterns. The most valuable memory for a tutoring system is not “which question the user got wrong,” but “why the user keeps making the same mistake.”
For example:
Third, it lacks cross-module consolidation of learning events. DeepTutor has modules such as Chat, Deep Solve, Quiz, Notebook, Book, Research, and TutorBot. Each module generates learning signals, but these signals need to enter memory in a unified way.
For example:
All of these should be treated as learning evidence, rather than logs scattered across different features.
Fourth, it lacks explainable and correctable memory. If the system judges that “the user does not understand a concept,” the user should be able to see the evidence and modify or delete incorrect judgments.
For example:
3. Goals
The goal of this RFC is to upgrade DeepTutor memory from “remembering user information” to “supporting personalized teaching decisions.”
Specific goals include:
4. Recommend Memory Archteciture for DeepTutor
DeepTutor’s memory can be divided into six layers.
4.1 Profile Memory: User Profile
This layer records relatively stable information about the user.
It includes learning goals, knowledge background, preferred explanation style, language preference, learning pace, and current areas of focus.
DeepTutor already has basic capabilities in this area, which can be retained.
4.2 Summary Memory: Learning Summary
This layer records the user’s recent and long-term learning trajectory.
It includes what the user has recently learned, which tasks have been completed, which topics have been explored, and how the user’s understanding has changed over time.
4.3 Learner State Memory: Learning State
This is the new core layer.
It records the user’s state for each knowledge point.
For example:
This memory layer determines whether DeepTutor can truly provide personalized teaching.
4.4 Misconception Memory: Misconceptions and Error Patterns
This layer records mistakes that repeatedly appear for the user.
This memory layer allows the system to avoid giving generic explanations repeatedly and instead directly correct the key misconception.
4.5 Artifact Memory: Learning Asset Memory
This layer incorporates learning artifacts such as Notebook, Question Bank, Book, and Research notes into memory.
It aims to answer the following questions:
Artifacts are not merely files. They are learning evidence.
4.6 TutorBot Private Memory: TutorBot Private Memory
Each TutorBot can have its own memory.
Related Module
Other
Use Case
No response
Additional Context
No response