## Describe what this PR does / why we need it
**Current situation**:
The… current `DeepResearch` project lacks conversation memory, after a round of conversation, the model cannot remember what the user just said.
Based on the above background, this PR provides the following two feat implementations.
1. **Conversation Memory**: Supports recording user questions and the final agent's answers or reports.Support for the use of session memory in deepresearch multi-turn dialogues
2. **User Role Memory**: Learns users' potential characteristics in real time and continuously through multi-turn conversations by analyzing their chat content, and uses these characteristics to guide the agent to generate conversation content that is more suitable for the user.
## Describe how you did it
### Conversation Memory
The implementation of conversation memory is fairly conventional, still using Spring AI's `org.springframework.ai.chat.memory.MessageWindowChatMemory`, and then manually recording it at the nodes that need to be recorded throughout the DeepResearch project.
In this feat, conversation memories are primarily stored and retrieved in the `CoordinatorNode` and `ReporterNode`, corresponding to the agent accepting requests and returning responses, respectively.
### User Role Memory
#### Motivation
User role memory stems from user habits encountered during conversations with the agent—users often inform the agent of their professional background to guide the agent in producing responses more appropriate to their context.
For example, the agent's response will be significantly different if the user assumes they are a university student versus an elementary school student.
However, relying solely on the user to guide the agent is not intelligent enough.
**Our goal is to enable agents to proactively learn user role information from conversations and use this information to guide LLMs in real time and at the conversational level, while supporting adaptive evolution and updates of role information across multiple rounds of conversation.**
#### Architecture
User role memory uses a method similar to `mem0` for memory extract and updating.
Unlike mem0, conversation-level memory bootstrapping requires speed and real-time processing, focusing more on short-term memory rather than generating `knowledge graph` content such as entity-relationship structures. Therefore, in its implementation, role memory generally uses the same storage method as short-term memory, such as in-memory or Redis.
**User role memory is divided into two stages**:
1. extract
This defines a `ShortUserRoleExtractResult` entity, which is detailed below, including `ConversationAnalysis`, `IdentifiedRole`, and `CommunicationPreferences`. This attempts to describe the user's current role and communication preferences using these three methods.
```ts
interface ConversationAnalysis {
confidenceScore: number; // The confidence score ranges from 0 to 1
interactionCount: number; // The number of interactions in the current session
}
interface IdentifiedRole {
possibleIdentities : string[]; // List of possible identities, like "software_engineer", "housewife", etc.
primaryCharacteristics: string[]; // Main character feature tags
evidenceSummary: string[]; // Summary of identification basis
confidenceLevel: 'LOW' | 'MEDIUM' | 'MEDIUM_HIGH' | 'HIGH'; // Confidence level
}
interface CommunicationPreferences {
detailLevel: 'CONCISE' | 'BALANCE' | 'COMPREHENSIVE'; // Detail preference level
contentDepth: 'OVERVIEW' | 'PRACTICAL' | 'CONCEPTUAL'; // Content depth
responseFormat: 'CONCISE' | 'DETAILED' | 'STRUCTURED_WITH_EXAMPLES'; // Preference response format
}
interface ShortUserRoleExtractResult {
conversationAnalysisInfo: ConversationAnalysis;
identifiedRole: IdentifiedRole;
communicationPreferences: CommunicationPreferences;
userOverview: string; // To describe User information in one sentence based on identifiedRole and communicationPreferences, the occupation information in possibleIdentities must be used
}
```
For each user's `question`, `ShortMemoryAgent` will refer to `historical user questions` to extract user role memories.
**Q: For extraction functionality, why do we only need to consider the user's question and not provide the LLM answer within the context?**
**A**: This stems from a phenomenon where, during interpersonal communication, humans can observe the speaker's words to infer their approximate identity and communication habits. However, what the other person says during communication doesn't determine my own role or background; rather, what I say is most important in defining my user identity.
2. update
Here are three important memories.
- `Current Extract Memory`：The current user role memory generated in the `extract phase` is based on current user questions and historical user questions.
- `Previous Extract Memory`：The latest user role memory. In a single conversation, there is only one up-to-date user role memory, because role memory should evolve continuously with each user's questions.
- `History Extract Track`：Historical User Role Memory Evolution Trajectory
During the update phase, this section transmits the current user role memory (`Current Extract Memory`) and the most recent user role memory (`Previous Extract Memory`) extracted in the current round, as well as the historical user role memory track (`History Extract Track`). If the extracted current user role memory and the most recent user role memory exceed a set threshold, an `UPDATE` operation is triggered, during which similar words and features are merged. If the two are significantly different, the History Extract Track is referenced to determine whether to perform an `UPDATE`. If they are still different, a `NONE CHANGE` is triggered, and the most recent user role memory is returned.
#### Logical Overview
The extraction and update logic is implemented in `ShortUserRoleMemoryNode`. This logic is designed to make the user role memory mechanism more suitable for multi-turn conversation scenarios.
<img width="8897" height="757" alt="会话引导机制-横板" src="https://github.com/user-attachments/assets/38913d15-b360-42f5-bdc0-ae6185139d62" />
### Case Study
**Conversation memory (No plan)**：
<img width="805" height="858" alt="会话记忆-未触发工具调用" src="https://github.com/user-attachments/assets/fbf068f5-3799-45dd-bf9a-bf690a4033ef" />
```json
{
"code": 200,
"status": "success",
"message": "",
"data": [
{
"messageType": "USER",
"metadata": {
"messageType": "USER"
},
"media": [
],
"text": "我想做红烧肉给家人吃"
},
{
"messageType": "ASSISTANT",
"metadata": {
"finishReason": "STOP",
"search_info": "",
"id": "8be26c58-0881-4601-9dba-b7bc0921f162",
"role": "ASSISTANT",
"messageType": "ASSISTANT",
"reasoningContent": ""
},
"toolCalls": [
],
"media": [
],
"text": "红烧肉是一道非常受欢迎的家常菜，做法也相对简单。下面是一个基础版本的红烧肉食谱供您参考：\n\n### 材料\n- 五花肉 500克\n- 生姜几片\n- 葱 2根\n- 八角 2颗\n- 老抽、生抽各适量\n- 糖（最好是冰糖）3大勺\n- 清水足够覆盖肉块\n- 料酒少许\n\n### 做法\n1. **准备材料**：将五花肉切成约2厘米见方的小块；生姜切片；葱打结。\n2. **焯水**：把切好的五花肉放入冷水中，加入少量料酒和几片生姜，大火煮开后撇去浮沫，捞出洗净备用。\n3. **炒糖色**：锅中放少量油，加入糖小火慢慢加热至糖融化并开始变色成琥珀色时，迅速加入焯好水的五花肉快速翻炒均匀，使肉块表面裹上糖色。\n4. **调味炖煮**：向锅中加入足够的热水（没过肉），再加入剩余的生姜片、葱结、八角以及老抽调色调味，大火烧开后转小火慢炖约1小时左右或直到肉质软烂。\n5. **收汁**：待肉快熟时开大火收汁，根据个人口味调整盐分，并且可以适当增加一些生抽提鲜。\n6. **完成**：最后撒上葱花即可出锅享用啦！\n\n希望这个简单的步骤能帮助到你，给家人带来一顿美味的红烧肉！如果需要更详细的步骤或者其他建议，请随时告诉我。"
},
{
"messageType": "USER",
"metadata": {
"messageType": "USER"
},
"media": [
],
"text": "有老人和小孩，可以调整下食谱吗"
},
{
"messageType": "ASSISTANT",
"metadata": {
"finishReason": "STOP",
"search_info": "",
"id": "1571fcda-d3a6-4c33-8111-8126fad74c96",
"role": "ASSISTANT",
"messageType": "ASSISTANT",
"reasoningContent": ""
},
"toolCalls": [
],
"media": [
],
"text": "当然可以调整，考虑到老人和小孩的口味偏好以及消化能力，这里有一些小建议来调整红烧肉的做法：\n\n1. **减少糖量**：因为老人和小孩可能对甜食比较敏感，你可以适当减少糖的用量。如果原本使用了3大勺糖，那么可以尝试减至2大勺或更少。\n\n2. **增加蔬菜**：在炖煮过程中加入一些容易消化且富含纤维素的蔬菜如胡萝卜、土豆等，不仅能让菜肴更加营养均衡，还能让口感更加丰富多样。\n\n3. **软烂易嚼**：确保五花肉炖得足够软烂，这样更容易被老人和孩子咀嚼吞咽。烹饪时间可以适当延长至1小时半甚至更长，直到肉质非常柔软为止。\n\n4. **调味清淡**：除了减少糖外，还可以稍微减少酱油（尤其是老抽）的使用量以降低咸度；同时避免使用过多辛辣调料，保持整体味道温和。\n\n5. **去皮处理**：对于五花肉中的肥肉部分，可以选择去除表皮或者只保留少量脂肪层，这样既能保证菜品的风味又不会过于油腻。\n\n按照以上几点调整后的红烧肉应该会更适合全家人的口味哦！希望这些建议对你有所帮助。"
}
]
}
```
**User Role Memory (Plan)**：
History User Role Extraction Track - Role Information Evolves Based on User Chat
<img width="5632" height="1936" alt="用户角色记忆进化轨迹" src="https://github.com/user-attachments/assets/994ee359-8b11-4e20-8127-c59e7dd5d80f" />
Source json
```json
{
"code": 200,
"status": "success",
"message": "",
"data": [
{
"messageType": "SYSTEM",
"metadata": {
"messageType": "SYSTEM"
},
"text": {
"userId": "MOCK_USER_ID",
"conversationId": "7ad1d39a-d8b5-41d0-aaa1-c6442fb4fb83",
"userQuery": "研究下回锅肉的做法",
"conversationAnalysis": {
"confidenceScore": 0.7,
"interactionCount": 1
},
"identifiedRole": {
"possibleIdentities": [
"家庭主妇",
"烹饪爱好者"
],
"primaryCharacteristics": [
"注重实际操作",
"关注具体步骤"
],
"evidenceSummary": [
"询问了回锅肉的做法"
],
"confidenceLevel": "MEDIUM"
},
"communicationPreferences": {
"detailLevel": "CONCISE",
"contentDepth": "PRACTICAL",
"responseFormat": "CONCISE"
},
"userOverview": "一位注重实际操作的家庭主妇或烹饪爱好者，倾向于简洁明了的步骤说明",
"creatTime": "2025-12-01 18:18:19",
"updateTime": ""
}
},
{
"messageType": "SYSTEM",
"metadata": {
"messageType": "SYSTEM"
},
"text": {
"userId": "MOCK_USER_ID",
"conversationId": "7ad1d39a-d8b5-41d0-aaa1-c6442fb4fb83",
"userQuery": "改进一下口味，让他更加适合于老人和小孩",
"conversationAnalysis": {
"confidenceScore": 0.8,
"interactionCount": 2
},
"identifiedRole": {
"possibleIdentities": [
"家庭主妇",
"烹饪爱好者"
],
"primaryCharacteristics": [
"注重健康饮食",
"关注家庭成员需求"
],
"evidenceSummary": [
"要求改进菜的口味以适合老人和小孩",
"之前询问了回锅肉的做法"
],
"confidenceLevel": "HIGH"
},
"communicationPreferences": {
"detailLevel": "BALANCE",
"contentDepth": "PRACTICAL",
"responseFormat": "STRUCTURED_WITH_EXAMPLES"
},
"userOverview": "一位注重健康饮食的家庭主妇或烹饪爱好者，关注家庭成员的需求，希望获得适合老人和小孩的实用食谱建议",
"creatTime": "2025-12-01 18:24:35",
"updateTime": "2025-12-01 18:24:59"
}
},
{
"messageType": "SYSTEM",
"metadata": {
"messageType": "SYSTEM"
},
"text": {
"userId": "MOCK_USER_ID",
"conversationId": "7ad1d39a-d8b5-41d0-aaa1-c6442fb4fb83",
"userQuery": "我想在进一步研究一下适合于家庭成员的回锅肉做法，详细一点",
"conversationAnalysis": {
"confidenceScore": 0.8,
"interactionCount": 3
},
"identifiedRole": {
"possibleIdentities": [
"家庭主妇",
"烹饪爱好者"
],
"primaryCharacteristics": [
"注重实际操作",
"关注细节",
"注重健康饮食",
"关注家庭成员需求"
],
"evidenceSummary": [
"询问了回锅肉的详细做法",
"之前询问过回锅肉的做法",
"要求改进菜的口味以适合老人和小孩"
],
"confidenceLevel": "HIGH"
},
"communicationPreferences": {
"detailLevel": "COMPREHENSIVE",
"contentDepth": "PRACTICAL",
"responseFormat": "STRUCTURED_WITH_EXAMPLES"
},
"userOverview": "一位注重实际操作、细节、健康饮食和家庭成员需求的家庭主妇或烹饪爱好者，希望获得详细的回锅肉做法说明以及适合老人和小孩的改良建议",
"creatTime": "2025-12-01 18:26:12",
"updateTime": "2025-12-01 18:26:27"
}
}
]
}
```
Related user conversastion memory
```json
{
"code": 200,
"status": "success",
"message": "",
"data": [
{
"messageType": "USER",
"metadata": {
"messageType": "USER"
},
"media": [
],
"text": "研究下回锅肉的做法"
},
{
"messageType": "ASSISTANT",
"metadata": {
"messageType": "ASSISTANT"
},
"toolCalls": [
],
"media": [
],
"text": "# Key Points\n- 回锅肉是一道经典川菜，核心步骤包括煮肉、煸炒和调味。\n- 煮肉时需冷水下锅，加入料酒和葱段，煮至筷子能轻松插入后切片。\n- 煸炒时要将五花肉的油脂充分炒出，搭配蒜米或花椒增加香味。\n- 调味料的选择和火候控制是回锅肉香而不腻的关键。\n- 不同做法在细节上有所差异，但基本流程一致。\n\n---\n\n# Overview\n回锅肉作为川菜中的经典代表，以其鲜香不腻、口感丰富的特点深受喜爱。这道菜的核心在于通过煮、炒两步工艺处理五花肉，并结合调味料使其风味浓郁。本报告旨在梳理回锅肉的烹饪步骤，分析其关键环节，并提供简洁明了的操作指南，帮助家庭主妇或烹饪爱好者快速掌握制作技巧。\n\n---\n\n# Detailed Analysis\n\n## 1. 煮肉：奠定基础口感\n煮肉是回锅肉的第一步，也是确保肉质鲜嫩的重要环节。根据研究发现，以下步骤至关重要：\n- **冷水下锅**：冷水煮肉可以避免肉质因骤热而收缩，保持肉质柔软。\n- **去腥增香**：加入料酒和葱段可有效去除腥味，同时增添香气。\n- **煮至适中**：煮至表皮变白且筷子能轻松插入即可，避免过度煮烂影响后续煸炒。\n\n| 步骤 | 操作方法 | 目的 |\n|----------------|-------------------------------------|--------------------|\n| 冷水下锅 | 将五花肉与冷水一同加热 | 避免肉质收缩 |\n| 加入调料 | 放入料酒、葱段 | 去腥增香 |\n| 控制时间 | 煮至筷子能轻松插入 | 确保肉质适中 |\n\n案例参考：某家常做法建议煮好后静置半小时再切片，以进一步锁住肉汁（来源见参考资料）。\n\n---\n\n## 2. 煸炒：提升香气与口感\n煸炒是回锅肉的灵魂步骤，通过高温将五花肉中的油脂逼出，使成品香而不腻。具体操作如下：\n- **热锅冷油**：使用菜籽油时无需等油热即可下锅，中小火煸炒更佳。\n- **灯盏窝形状**：当肉片卷曲成“灯盏窝”状时，说明水分已被炒干，此时肉质最佳。\n- **辅料搭配**：加入蒜米或花椒等辅料，进一步提升香气。\n\n| 关键点 | 具体操作 | 效果 |\n|----------------|------------------------------------|--------------------|\n| 灯盏窝状态 | 中小火煸炒至肉片卷曲 | 提升口感层次 |\n| 辅料选择 | 蒜米、花椒 | 增加香气 |\n| 油脂处理 | 充分炒出肥肉中的油脂 | 减少油腻感 |\n\n案例参考：某餐厅级做法强调使用菜籽油并先下肉片，以保留传统风味。\n\n---\n\n## 3. 调味：赋予独特风味\n调味是回锅肉最后一步，决定了整道菜的风味特色。常见的调味料包括豆瓣酱、甜面酱、酱油等，这些调料不仅提味，还能让肉片更加红亮诱人。\n- **豆瓣酱**：为回锅肉提供咸香微辣的基础味道。\n- **甜面酱**：增加一丝甜味，平衡整体口感。\n- **翻炒均匀**：调味料需与肉片充分混合，确保每一片肉都入味。\n\n| 调味料 | 功能 | 使用建议 |\n|---------------|----------------------------------|--------------------|\n| 豆瓣酱 | 提供咸香微辣 | 适量即可 |\n| 甜面酱 | 增加甜味 | 根据个人口味调整 |\n| 酱油 | 上色 | 最后加入 |\n\n---\n\n## 4. 细节优化：个性化调整\n不同地区和个人习惯可能会对回锅肉的做法进行细微调整，例如：\n- **火候控制**：全程中小火更适合新手操作，避免烧焦。\n- **辅料变化**：青椒、洋葱等蔬菜可作为配菜，丰富营养和口感。\n- **提前准备**：煮好的五花肉可冷藏保存，方便随时取用。\n\n---\n\n# Survey Note (Optional)\n本报告综合了多篇关于回锅肉的文章和教程，涵盖了从家常到餐厅级别的多种做法。由于部分资料未标明具体数据，因此在实际操作中可能需要根据个人经验进行微调。\n\n---\n\n# Key Citations\n- [回锅肉的做法_经典川菜，下饭好菜](https://home.meishichina.com/recipe-130687.html) \n- [如何炒出餐廳等級的回鍋肉？正確作法別漏掉！](https://blog.icook.tw/posts/145604) \n- [回锅肉的家常做法，这样制作香浓不腻，美味下饭](http://www.360doc.com/content/22/1203/23/1339386_1058733874.shtml) \n- [薇式回锅肉【中餐厅】的做法步骤](https://m.xiachufang.com/recipe/104046904/) \n- [今天研究有香草味的回锅肉](https://www.facebook.com/ljloo68/posts/2797129930603713/)"
},
{
"messageType": "USER",
"metadata": {
"messageType": "USER"
},
"media": [
],
"text": "改进一下口味，让他更加适合于老人和小孩"
},
{
"messageType": "ASSISTANT",
"metadata": {
"finishReason": "STOP",
"search_info": "",
"role": "ASSISTANT",
"id": "7a4b90b7-faf6-4ca8-a0d5-40b93ffcb2c0",
"messageType": "ASSISTANT",
"reasoningContent": ""
},
"toolCalls": [
],
"media": [
],
"text": "为了使回锅肉更适合老人和小孩食用，我们可以从以下几个方面进行调整：\n\n1. **减少辛辣**：老人和小孩的肠胃可能对辛辣食物比较敏感，因此可以减少或省略豆瓣酱的使用，或者选择辣度较低的豆瓣酱。如果完全不加豆瓣酱，可以通过增加甜面酱的比例来弥补风味。\n\n2. **软化肉质**：确保五花肉煮得足够软烂，这样更容易咀嚼和消化。可以在煮肉时延长烹饪时间，并且在切片前让肉稍微冷却一下，这样更容易切成薄片而不散开。\n\n3. **健康调味**：使用低钠酱油和其他调味品，以减少盐分摄入；同时控制糖的用量，避免过甜。可以考虑加入一些天然食材如姜、蒜等来提味，而不是过多依赖加工调料。\n\n4. **增加蔬菜**：添加适量的蔬菜不仅可以丰富营养成分，还能改善口感。例如，胡萝卜、青椒（去籽）、洋葱都是不错的选择。这些蔬菜不仅颜色鲜艳吸引孩子，而且富含维生素和纤维素，有助于促进消化。\n\n5. **温和调味**：采用更加温和的调味方式，比如使用少量的生抽、老抽调色，以及适量的白糖提鲜，保持整体口味清新自然。\n\n### 改良版回锅肉食谱\n\n#### 材料\n- 五花肉 300克\n- 胡萝卜 1根\n- 青椒 1个（可选）\n- 洋葱 半个\n- 生姜 几片\n- 大蒜 几瓣\n- 甜面酱 1大勺\n- 低钠酱油 1小勺\n- 白糖 半小勺\n- 料酒 适量\n- 葱段 适量\n- 清水 适量\n- 菜籽油 适量\n\n#### 做法\n1. **准备材料**：将五花肉洗净后放入冷水中，加入几片生姜和葱段，大火烧开后撇去浮沫，转中小火慢炖约30分钟至肉质变软但不散。\n \n2. **切片处理**：待肉凉透后切成薄片备用；胡萝卜、青椒、洋葱分别洗净切丝。\n\n3. **炒制过程**：\n - 热锅冷油下入五花肉片，用中小火慢慢煸炒至表面微黄并卷曲成“灯盏窝”状。\n - 加入切好的姜蒜末翻炒出香味。\n - 接着依次加入胡萝卜丝、青椒丝及洋葱丝快速翻炒均匀。\n - 最后调入甜面酱、低钠酱油、白糖以及少许料酒继续翻炒至上色均匀即可出锅装盘。\n\n通过以上改良措施，可以使这道传统川菜变得更加适合家庭成员中不同年龄段的人群享用，既保留了原有的美味特色又兼顾了健康需求。希望这个建议对你有所帮助！"
},
{
"messageType": "USER",
"metadata": {
"messageType": "USER"
},
"media": [
],
"text": "我想在进一步研究一下适合于家庭成员的回锅肉做法，详细一点"
},
{
"messageType": "ASSISTANT",
"metadata": {
"messageType": "ASSISTANT"
},
"toolCalls": [
],
"media": [
],
"text": "# Key Points\n- 回锅肉的核心步骤包括煮肉、煸炒和调味，使用五花肉、豆瓣酱、蒜苗等基础食材。\n- 适合老人和小孩的改良版回锅肉建议减少油腻感、降低辣度，并增加蔬菜搭配。\n- 健康调整包括选用瘦肉比例更高的部位、减少油脂用量、搭配易消化的蔬菜。\n- 烹饪时注意火候控制，避免过焦或过油，确保肉质嫩滑且易于咀嚼。\n\n---\n\n# Overview\n回锅肉是一道经典的川菜，以其香辣浓郁的味道深受喜爱。然而，传统的回锅肉可能过于油腻或辛辣，不适合老人和小孩食用。本报告旨在提供一份详细的家庭版回锅肉制作方法，并针对老人和小孩的需求提出改良建议，以满足不同家庭成员的口味与健康需求。\n\n---\n\n# Detailed Analysis\n\n## 1. 回锅肉的基本做法\n回锅肉的传统制作过程分为三个主要步骤：煮肉、煸炒和调味。以下是详细的步骤说明：\n\n### (1) 准备食材\n- **主料**：选择肥瘦相间的五花肉（约300克），确保口感层次丰富。\n- **辅料**：青椒、蒜苗、姜片、蒜末等。\n- **调料**：郫县豆瓣酱、甜面酱、酱油、料酒、白糖、盐。\n\n### (2) 煮肉\n- 将五花肉冷水下锅，加入葱段、姜片和料酒去腥，煮至筷子能轻松插入（约15-20分钟）。\n- 捞出后冷却，切成约2元硬币厚的薄片，便于后续煸炒。\n\n### (3) 煸炒\n- 热锅冷油，将切好的肉片放入锅中，用中小火煸炒至卷曲成“灯盏窝”状，逼出多余油脂。\n- 加入姜片、蒜末、豆瓣酱翻炒出红油，随后加入甜面酱、酱油等调味料。\n\n### (4) 调味与收尾\n- 加入青椒、蒜苗等配菜，快速翻炒均匀，最后根据个人口味调整盐分即可出锅。\n\n---\n\n## 2. 改良建议：适合老人和小孩的版本\n为了适应老人和小孩的饮食需求，可以从以下几个方面对传统回锅肉进行改良：\n\n### (1) 减少油腻感\n- **选用瘦肉比例更高的部位**：如猪里脊或鸡胸肉，减少脂肪含量。\n- **控制油脂用量**：煸炒时尽量逼出多余油脂，避免过油。\n- **使用健康油品**：如橄榄油或玉米油，代替传统的菜籽油，减少饱和脂肪摄入。\n\n### (2) 降低辣度\n- **减少辣椒和辣味调料**：如豆瓣酱和辣椒酱的用量，以适应小孩和老年人的口味。\n- **增加甜味**：适当加入少量糖或甜面酱，让味道更柔和。\n- **使用低盐酱油**：减少钠摄入，更适合老年人健康饮食。\n\n### (3) 增加营养搭配\n- **搭配蔬菜**：加入青椒、胡萝卜、洋葱等蔬菜，提升维生素和纤维素摄入。\n- **搭配主食**：如米饭或面条，有助于消化吸收。\n\n### (4) 切片更薄\n- 将五花肉切得更薄一些，使其更容易咀嚼和消化，尤其适合老人和小孩。\n\n---\n\n## 3. 其他注意事项\n- **提前准备**：煮好的五花肉可以冷藏保存，方便下次快速使用。\n- **火候控制**：全程使用中小火，避免烧焦或粘锅。\n- **卫生安全**：注意食材新鲜度，尤其是肉类，避免变质引发健康问题。\n\n---\n\n# Survey Note\n本报告综合了多个来源的视频教程、文字说明以及用户评论，结合实际操作经验，提出了适合家庭制作的回锅肉详细步骤及改良建议。由于部分数据来源于非专业渠道，可能存在一定主观性。\n\n---\n\n# Key Citations\n- [回锅肉这样做，简单又好吃 - TikTok](https://www.tiktok.com/@user6963761036556/video/7577379774792879378) \n- [轻松做出家常回锅肉,美食,美食教学,好看视频](https://haokan.baidu.com/v?pd=wisenatural&vid=11344966261624000822) \n- [祭拜肉品這樣做！簡單又下飯！冰箱剩食也能大翻身【回鍋肉】](https://www.youtube.com/watch?v=wMtcsfJXgjI) \n- [老公最爱回锅肉，简单做一锅，川味十足，家里也能轻松复制！](https://haokan.baidu.com/v?pd=wisenatural&vid=5462758488125479642) \n- [在家做回锅肉，轻松掌握家常美味！](https://haokan.baidu.com/v?pd=wisenatural&vid=6178495915616329954)"
}
]
}
```
**User role memory guidance effect**
<img width="3002" height="1588" alt="引导效果png" src="https://github.com/user-attachments/assets/90cc3fde-0ad4-4944-b2fa-1744573a540b" />
### Short Term Memory Properties
```yaml
spring:
ai:
alibaba:
deepresearch:
# short-term memory configuration
short-term-memory:
enabled: true
# conversation memory configuration
conversation-memory:
# maximum Reservation conversation Length
max-messages: 100
# user role memory configuration
user-role-memory:
# scope of User Role Short-Term Memory Guidance Model Generation
guide-scope: every
# The similarity threshold for determining when a memory update is needed ranges from 0 to 1, with a default of 0.8.
update-similarity-threshold : 0.8
# Refer to the number of questions asked by users in the past
history-user-messages-num: 10
# memory-type
memory-type: in-memory
```
### Http EndPoint
Get conversation history messages
- GET http://localhost:8080/api/user/memory/conversation
Get user roles extract short-term memory track
- GET http://localhost:8080/api/user/memory/track
Get latest user role memory
- GET http://localhost:8080/api/user/memory/latest
Delete user role memory
- POST http://localhost:8080/api/user/memory/delete?session_id=__default__
洛卡卡了 08-18 17:111楼
后仰跳投 08-18 17:572楼
qyy 08-18 18:595楼
灿烂甜菜 08-18 19:106楼
甘尼克斯 08-18 23:2110楼