memory（记忆）

Memory Management

Efficient context and knowledge management.

Instructions

1. Working Memory Model

┌─────────────────────────────────────────┐
│           WORKING MEMORY                │
├─────────────────────────────────────────┤
│ • Current task goal                     │
│ • Relevant file contents                │
│ • Recent decisions                      │
│ • Active constraints                    │
└─────────────────────────────────────────┘
          ↑ Load        ↓ Store
┌─────────────────────────────────────────┐
│          LONG-TERM MEMORY               │
├─────────────────────────────────────────┤
│ • Project structure                     │
│ • User preferences                      │
│ • Past solutions                        │
│ • Domain knowledge                      │
└─────────────────────────────────────────┘

2. Context Prioritization

Order of importance for context:

3. Information Retention

## What to Remember

✅ Keep in context:
- Current task objective
- File being modified
- Type definitions in use
- Recent error messages
- User preferences

❌ Safe to forget:
- Already processed files
- Resolved errors
- Intermediate calculations
- Verbose logs

4. Context Summarization

When context grows too large:

## Summarization Rules

1. **Files**: Keep imports, types, key functions
2. **Errors**: Keep message, remove stack trace
3. **Logs**: Keep last 10 lines
4. **History**: Keep decisions, remove process

### Example

Before (verbose):
"I looked at file A, then file B, noticed pattern X,
then explored file C, found issue Y, traced it to..."

After (summarized):
"Analyzed A, B, C. Found: pattern X, issue Y in C."

5. Session State Pattern

// Conceptual session state
interface SessionMemory {
  // Always retain
  task: {
    goal: string;
    status: 'planning' | 'executing' | 'verifying';
    progress: number;
  };

  // Retain while relevant
  context: {
    activeFiles: string[];
    recentDecisions: string[];
    constraints: string[];
  };

  // Summarize or forget
  history: {
    summary: string;
    keyInsights: string[];
  };
}

6. Knowledge Retrieval

## Before Starting New Task

1. Check: Have I seen this before?
2. Recall: What approach worked?
3. Apply: Use proven patterns
4. Adapt: Modify for current context

7. Memory Hygiene

## Per-Turn Cleanup

After completing a step:

1. ✅ Task still relevant? Keep
2. ❓ Might need later? Summarize
3. ❌ No longer needed? Forget

## End of Task

1. Extract learnings
2. Update knowledge base
3. Clear working memory

8. Context Window Management

## Token Budget Allocation

| Category | Budget |
|----------|--------|
| System prompt | 10% |
| Task context | 30% |
| Active code | 40% |
| Conversation | 20% |

## When Near Limit

1. Summarize conversation history
2. Remove resolved issues
3. Keep only relevant code sections
4. Preserve critical context

References

• Working Memory in LLMs
• Context Compression