home / skills / zephyrwang6 / myskill / mem-query

mem-query skill

safe

This skill retrieves and synthesizes user memories from multiple layers, citing sources to answer personal history, habits, and preferences.

npx playbooks add skill zephyrwang6/myskill --skill mem-query

Review the files below or copy the command above to add this skill to your agents.

Files (1)

SKILL.md

3.4 KB

---
name: mem-query
description: AI个人记忆系统的记忆查询功能。检索各层级记忆文件，综合多层级信息回答用户问题。使用场景：(1) 用户问"我的记忆中关于XXX"时；(2) 用户询问自己的习惯/偏好/价值观时；(3) 需要基于用户历史提供建议时。该skill会自动检索L1-L4各层级，引用来源，给出基于记忆的个性化回答。
---

# 记忆查询

## 快速开始

当用户询问关于自己的历史时：

1. **读取INDEX.md** - 获取记忆地图
2. **确定检索层级** - 根据问题确定相关层级
3. **Grep搜索** - 使用关键词搜索
4. **读取文件** - 读取匹配的文件
5. **综合回答** - 结合多层级信息回答

## 检索协议

### 标准流程

```
用户提问
  ↓
读取 AI_MEMORY/INDEX.md
  ↓
识别问题类型 → 确定相关层级
  ↓
Grep关键词搜索 → AI_MEMORY/
  ↓
读取匹配文件
  ↓
综合多层级信息
  ↓
用L4价值观校准回答
  ↓
生成回复（引用来源）
```

### 检索优先级

1. **L4核心层** - 价值观校准（必须读取）
2. **L3认知层** - 思维框架（如涉及决策）
3. **L2行为层** - 行为模式（如涉及具体做法）
4. **L1情境层** - 近期上下文（如涉及当前项目）

### Grep搜索策略

```bash
# 示例1：用户问工作习惯
Grep "工作习惯|工作流|处理.*任务" AI_MEMORY/L2_行为层/

# 示例2：用户问如何决策
Grep "决策|选择|判断" AI_MEMORY/L3_认知层/

# 示例3：用户问价值观相关
Grep "价值观|重要|原则" AI_MEMORY/L4_核心层/
```

## 回答格式

**标准回答模板**：

```
根据你的{层级}记录（{文件名}）：

{具体内容}

{来源引用}
```

**示例**：

```
根据你的工作习惯记录（L2_行为层/工作习惯.md）：

你处理复杂任务时通常会：
1. 先用思维导图拆解（出现频率：5次）
2. 按优先级排序（决策框架：MECE原则）
3. 分步执行，每步验证

你希望我按这个方式来处理吗？
```

## 常见查询类型

### 查询工作习惯

```
Q: "我通常如何处理复杂任务？"

检索：
- Grep "复杂任务" L2_行为层/
- 读取 L2_行为层/工作习惯.md
- 读取 L3_认知层/决策框架.md
```

### 查询价值观

```
Q: "我最看重什么？"

检索：
- 读取 L4_核心层/核心价值观.md（必须）
- 不需要搜索，直接读取
```

### 查询近期事件

```
Q: "我上周做了什么？"

检索：
- 读取 L1_情境层/{年}-{月}.md
- 提取相关日期范围的记录
```

### 查询偏好

```
Q: "我喜欢用什么工具？"

检索：
- Grep "工具|偏好" L2_行为层/
- 读取 L2_行为层/工具偏好.md
```

## 错误处理

### 找不到相关记忆

```
我在你的记忆中没找到相关记录。

这可能意味着：
1. 这是新情况，还没形成模式
2. 之前没记录

要记录这次对话到L1情境层吗？
```

### 记忆冲突

```
⚠️ 发现记忆冲突：
- L2_行为层/工作习惯.md 记录：你喜欢用思维导图
- L1_情境层/2025-12-15.md 记录：你今天说"不想用思维导图"

可能原因：
1. 偏好发生了变化
2. 情境不同

请帮我确认：哪个记录更准确？
```

## 重要原则

1. **总是引用来源** - 告诉用户信息来自哪个文件
2. **用L4校准** - 所有建议都要符合用户的价值观
3. **诚实无记忆** - 找不到就直说，不要编造
4. **区分层级** - 明确说明这是行为模式还是价值观

Overview

This skill provides a memory-query function for a personal AI memory system. It automatically retrieves and synthesizes records across hierarchical memory layers (L1–L4) to answer questions about your history, habits, preferences, and values. Responses include explicit source citations and are calibrated by your core values.

How this skill works

When you ask about something in your memory, the skill reads the memory index to map relevant files, selects which layers to search, greps matching entries, and reads the matched files. It combines information from L1 (recent context) through L4 (core values), applies L4-based calibration to any recommendations, and returns a concise, sourced answer.

When to use it

Ask "What do I remember about X?" to get an integrated summary from past records.
Request your habits, preferences, or recurring behaviors to guide personalized actions.
Ask for advice that should reflect your historical choices or stated values.
Check recent activity or events from a specific date or timeframe.
Verify or resolve apparent conflicts between different memory entries.

Best practices

Always include the memory index: the skill reads AI_MEMORY/INDEX.md first to find relevant files.
Prioritize L4 reads: core values are always loaded and used to calibrate recommendations.
Cite sources inline: every factual claim shows which memory file it came from.
Be explicit when memory is missing: do not fabricate—offer to record new context to L1.
Flag conflicts and prompt for clarification when different layers disagree.

Example use cases

You ask "How do I usually handle complex projects?" and get a synthesis from L2 behavior and L3 decision-framework notes with file citations.
You ask "What do I value most?" and the skill returns L4 core value statements directly from the core-values file.
You ask "What did I do last week?" and receive entries extracted from L1 daily logs for that date range.
You ask "Which tools do I prefer?" and get a list drawn from L2 tool-preferences with frequency/context notes.
You report contradictory entries and the skill shows both records, identifies the conflict, and asks which is current.

FAQ

What if no memory matches my question?

I will tell you I found no relevant records and offer to log this new information into L1 context so future queries can use it.

How do you handle conflicting records?

I surface both entries with file references, explain possible reasons (change over time or situational difference), and ask you to confirm which is accurate.