Memory-reme - ReMe Memory Management

A memory management system powered by ReMe that provides persistent cross-session memory, automatic user preference application, and intelligent context compression.

When to Use This Skill

Activate this skill when:

• User asks you to remember something ("记住这个", "别忘了", "下次注意")
• User provides feedback on your behavior ("你总是忘记", "为什么又这样")
• User refers to past information ("之前说过", "上次怎么做的")
• User asks about your preferences or settings
• User wants to prevent repeated mistakes
• Long conversations where context might overflow

Core Concepts

Three-Level Memory

1. Long-term Memory (MEMORY.md)

• User preferences and rules
• Persistent across all sessions
• Updated manually or through learning

2. Daily Memory (memory/YYYY-MM-DD.md)

• Session summaries
• Important events and decisions
• Auto-generated at session end

3. In-Memory Context

• Current conversation state
• Compressed when approaching limits
• Temporary, session-bound

Memory Types

Quick Start

Installation (One-time setup)

pip install reme-ai

Session Initialization

At the start of EVERY session:

1. Initialize ReMe
2. Retrieve user preferences
3. Apply to current context

# Initialize
from reme.reme_light import ReMeLight
reme = ReMeLight(working_dir=".reme", language="zh")
await reme.start()

# Retrieve preferences
prefs = await reme.memory_search(
    query="用户偏好 文件发送",
    max_results=5
)

# Apply
if prefs and "必须发送" in prefs[0]['content']:
    auto_send_files = True

Workflow

Phase 1: Session Start (0-5s)

┌─────────────────────────────────┐
│  1. Initialize ReMe            │
│  2. Load MEMORY.md            │
│  3. Search for user prefs     │
│  4. Apply to current context   │
└─────────────────────────────────┘

Action:

python3 C:\path\to\memory-reme\scripts\init_reme.py

Expected Output:

✓ ReMe initialized
📖 Retrieved 3 preferences
  - User prefers concise code
  - Files must be sent automatically
  - Prefer markdown over plain text
✓ Preferences applied

Phase 2: During Session

Check before actions:

1. Before generating files:
• Search for file handling preferences
• Apply formatting preferences

1. Before using tools:
• Search for tool-specific preferences
• Apply timeout/retry settings

1. User feedback:
• Extract new rules
• Add to MEMORY.md

Example:

User: "你怎么总是忘记发送文件？记住，生成文件后必须直接发送！"

Action:

# Learn from feedback
await reme.add_memory(
    memory_content="用户偏好：生成文件后必须使用message工具直接发送文件，不接受链接地址。原因：用户需要直观可见的内容。",
    user_name="阿伟",
    memory_type="personal"
)

Phase 3: Session End

┌─────────────────────────────────┐
│  1. Extract key events        │
│  2. Generate summary          │
│  3. Write to memory/          │
│  4. Update MEMORY.md         │
│  5. Cleanup tool results       │
│  6. Close ReMe               │
└─────────────────────────────────┘

Action:

python3 C:\path\to\memory-reme\scripts\save_summary.py

Output:

💾 Summary saved to memory/2026-03-06.md
✓ MEMORY.md updated
✓ Tool results cleaned
✓ ReMe closed

Common Use Cases

Use Case 1: File Generation

Trigger: User requests a file to be created

Workflow:

1. Check for file preferences
2. Generate file with correct format
3. Send automatically if required
4. Learn if user corrects

Example:

📖 Retrieved: "Send files automatically"

User: 生成AI日报

✓ Generated: AI日报_2026-03-06.md
📤 Sending file...
✓ Sent successfully

Use Case 2: Code Style Preferences

Trigger: User asks to write code

Workflow:

1. Search for style preferences
2. Apply conventions
3. Format accordingly

Example:

📖 Retrieved: "Prefer concise, well-commented code"

User: 写个Python函数

✓ Applied: Concise style with docstrings

Use Case 3: Preventing Repeated Mistakes

Trigger: User corrects your behavior

Workflow:

1. Accept feedback
2. Extract rule
3. Add to memory
4. Verify next time

Example:

User: Why do you keep forgetting to send files?

🧠 Learning...
✓ Rule recorded: "Always send files automatically"
✓ Will apply next time

Use Case 4: Context Overflow

Trigger: Conversation approaches 70% of token limit

Workflow:

1. ReMe automatically triggers
2. Compresses history to summary
3. Keeps critical information
4. Continues conversation

Automatic - no action needed.

Search Patterns

Common Search Queries

Search Results Processing

Always:

1. Review returned memories
2. Filter by relevance and recency
3. Apply to current context
4. Document what was applied

Example:

results = await reme.memory_search(query="文件发送 偏好", max_results=3)

for i, result in enumerate(results, 1):
    print(f"{i}. {result['content']}")
    if "必须发送" in result['content']:
        self.auto_send_files = True

print(f"✓ Applied: auto_send_files = {self.auto_send_files}")

Memory File Structure

MEMORY.md

# MEMORY.md - Long-term Memory

## User Profile
- **Name**: 阿伟
- **Role**: 90后程序员、AI博主

## Preferences

### File Handling
- **Rule**: 生成文件后必须使用message工具直接发送
- **Reason**: 用户需要直观可见的内容
- **Status**: Active
- **Learned**: 2026-03-06

### Code Style
- **Rule**: 代码要简洁，有注释
- **Reason**: 便于维护和理解
- **Status**: Active
- **Learned**: 2026-03-05

## Tool Usage

### web_fetch
- **Timeout**: 30s
- **Retry**: 3 times
- **Reason**: 某些网站响应慢

### browser
- **Timeout**: 60s
- **Wait time**: 3s for page load
- **Reason**: 确保页面完全加载

memory/YYYY-MM-DD.md

# 2026-03-06 Session Summary

## Session 1 - AI News Aggregation

### User Request
"给我今天的AI资讯"

### Processing
- Scraped 8 sources
- Filtered 20+ articles
- Selected 14 items

### Output
- File: AI日报_2026-03-06.md
- Size: 3611 bytes
- Sent: ✓

### User Feedback
"你怎么总是忘记发送文件？记住，生成文件后必须直接发送！"

### Learning
✓ New rule: Auto-send files
✓ Updated MEMORY.md

---

## Session 2 - ReMe Integration

### User Request
"接入ReMe后工作流程是怎样的"

### Processing
- Analyzed ReMe documentation
- Designed workflow
- Created integration plan

### Output
- File: ReMe工作流程设计.md
- File: ReMe存在形式与影响.md
- Sent: ✓

### No User Feedback

### Learning
No new rules

Best Practices

1. Always Start Sessions with Memory Retrieval

Bad:

# Start without memory
user_request = get_user_input()
process_request(user_request)

Good:

# Start with memory
reme = await init_reme()
prefs = await reme.memory_search(query="用户偏好")
apply_preferences(prefs)
user_request = get_user_input()
process_request(user_request)

2. Learn from Every Correction

When user says "You forgot X":

1. Acknowledge immediately
2. Extract the rule
3. Add to memory
4. Verify application

Example:

User: 你总是忘记发送文件！

Me: ✓ 已记住：生成文件后必须发送文件
   正在添加到 MEMORY.md...

Next file generation:
✓ File created
📤 Auto-sending...
✓ Sent

3. Be Specific in Memory Records

Bad:

- User prefers good code

Good:

- User prefers concise, well-commented Python code
  - Use docstrings for functions
  - Maximum 3 levels of nesting
  - Prefer list comprehensions over loops

4. Update Memory Regularly

Daily tasks:

• Review memory/ files
• Merge duplicate entries
• Remove outdated info
• Organize by category

Weekly tasks:

• Check for stale preferences
• Verify accuracy of tool settings
• Clean up old memory files

5. Use Semantic Search Effectively

Bad queries:

• "files"
• "code"
• "preferences"

Good queries:

• "文件发送 偏好 阿伟"
• "Python代码风格 简洁 注释"
• "工具设置 超时 重试"

Why: Specific queries return more relevant results.

Troubleshooting

Problem: Memory Not Retrieved

Symptoms:

• Preferences not applied
• Repeated mistakes
• Empty search results

Solutions:

1. Check if ReMe is initialized
2. Verify search query matches stored content
3. Check MEMORY.md exists and is not empty
4. Try broader search terms

# Debug search
results = await reme.memory_search(query="文件")
print(f"Found {len(results)} results")
for r in results:
    print(f"  - {r['content'][:50]}...")

Problem: Old Information Used

Symptoms:

• Outdated preferences applied
• Deprecated tool settings used

Solutions:

1. Add timestamp to memory entries
2. Sort results by time_created (reverse)
3. Manually update outdated entries in MEMORY.md
4. Consider expiration for time-sensitive rules

Problem: Memory File Too Large

Symptoms:

• MEMORY.md > 10KB
• Search slow
• Context bloat

Solutions:

1. Archive old entries to memory/archive/
2. Merge similar preferences
3. Remove redundant info
4. Use categories to organize

Integration with Existing Skills

Combining with docx skill

Workflow:

1. Search memory for docx preferences
2. Apply formatting rules
3. Generate document with docx skill
4. Check if auto-send required
5. Send if needed

Combining with coding-agent skill

Workflow:

1. Search memory for coding preferences
2. Apply style conventions
3. Generate code with coding-agent
4. Check for auto-review rules
5. Review if needed

Performance Considerations

Time Overhead

Space Usage

.reme/
├── MEMORY.md          ~10KB
├── memory/           ~150KB (30 days)
├── tool_result/       ~5MB (auto-cleanup)
└── .embeddings/       ~1MB

Total: ~6MB (1 month)

Advanced Features

Conditional Application

Only apply when relevant:

prefs = await reme.memory_search(query="文件发送")

if file_generated and prefs:
    # Apply file preferences
    if "必须发送" in prefs[0]['content']:
        await send_file(file_path)

Context-Aware Retrieval

Consider current task:

if task_type == "coding":
    query = "代码风格 Python"
elif task_type == "writing":
    query = "写作风格 简洁"
elif task_type == "file_generation":
    query = "文件发送 偏好"

Memory Cleanup

Automatic cleanup:

• Tool results expire after 7 days
• Embedding cache refreshed weekly
• Memory files archived monthly

Manual cleanup:

# Archive old sessions
mv memory/2026-01-*.md memory/archive/

# Compress large files
gzip MEMORY.md

See Also

• memory-structure.md - Detailed memory architecture
• best-practices.md - Advanced patterns
• common-prefs.md - Common preference examples

Summary

This skill enables persistent memory, automatic preference application, and intelligent context management. Use it to:

• ✓ Prevent repeated mistakes
• ✓ Remember user preferences
• ✓ Maintain context across sessions
• ✓ Learn from feedback
• ✓ Provide consistent behavior

Key principle: Memory is only useful when it's retrieved and applied. Always start sessions with memory retrieval, and verify application throughout the conversation.