dynamic-memory（动态记忆）

Dynamic Memory System 🧠

The intelligent memory system that thinks ahead.

Automatic importance detection. Heartbeat monitoring. Nightly review. Zero manual effort required.

Core Philosophy

Traditional memory systems are reactive — they wait for you to tell them what to remember.

Dynamic Memory is proactive — it thinks ahead, identifies what matters, and ensures nothing important slips through the cracks.

The Problem It Solves

❌ Traditional Memory:
   User: "Remember that I prefer dark mode"
   AI: "OK, I'll remember that"
   [User must explicitly ask AI to remember]

✅ Dynamic Memory:
   User: "I prefer dark mode"
   AI: [Automatically detects importance]
   AI: "Got it! I've noted your preference for dark mode."
   [Zero friction, automatic capture]

How It Works

Three-Layer Architecture

┌─────────────────────────────────────────────────────────┐
│            DYNAMIC MEMORY SYSTEM                        │
├─────────────────────────────────────────────────────────┤
│                                                          │
│  Layer 1: REAL-TIME CAPTURE                             │
│  ┌──────────────────────────────────────────────────┐  │
│  │  Every conversation → Analyze → Classify → Store  │  │
│  │  Triggers: End of each AI response               │  │
│  └──────────────────────────────────────────────────┘  │
│           ↓                                               │
│  Layer 2: HEARTBEAT MONITORING                          │
│  ┌──────────────────────────────────────────────────┐  │
│  │  Every 30 min → Re-evaluate → Merge → Dedupe     │  │
│  │  Boost importance of repeated mentions           │  │
│  └──────────────────────────────────────────────────┘  │
│           ↓                                               │
│  Layer 3: NIGHTLY REVIEW (9 PM)                         │
│  ┌──────────────────────────────────────────────────┐  │
│  │  Today's memories → Rank → Ask user → Archive    │  │
│  │  "Keep? Delete? Modify?"                          │  │
│  └──────────────────────────────────────────────────┘  │
│                                                          │
└─────────────────────────────────────────────────────────┘

When to Use This Skill

Automatic Activation (Recommended):

• Use this skill as a default enhancement for all conversations
• It runs silently in the background, only surfacing when important

Manual Activation:

• User explicitly asks to "remember something"
• User wants to review past memories
• User initiates a memory review session

AI Behavior Instructions

1. After Each Response (Real-Time Capture)

CRITICAL: Before considering your response complete, execute this mental checklist:

1. Did the user share:
   □ A preference? ("I like/dislike X")
   □ A decision? ("Let's use X instead of Y")
   □ A goal? ("I need to finish X by Friday")
   □ A correction? ("No, that's wrong, it should be X")
   □ Important context? ("By the way, X is important because...")

2. If ANY checkbox is marked:
   → Extract the key information
   → Assign importance score (0-1)
   → Categorize (preference/decision/goal/fact)
   → Store in temporary buffer
   → Briefly acknowledge: "Noted that [summary]"

Example:

User: "Let's use React for this project, not Vue."

AI Internal Process:
✓ Detected: Decision
✓ Importance: 0.95 (major technical decision)
✓ Category: decision
✓ Store: "Use React, not Vue, for this project"
✓ Acknowledge: "Got it — React it is. I've saved this decision."

AI Response: "Understood! I'll use React for this project. That's noted for future reference."

2. Importance Detection (Automatic)

Use these guidelines to auto-assign importance:

Detection Heuristics:

// Look for these linguistic markers
{
  "preference": ["prefer", "like", "dislike", "hate", "love", "favorite", "don't want"],
  "decision": ["let's use", "decided to", "going with", "choose", "selected"],
  "correction": ["no, actually", "wait, I meant", "that's wrong", "not X, but Y"],
  "goal": ["need to", "must", "by Friday", "deadline", "target", "goal is"],
  "importance": ["important", "critical", "essential", "must remember", "key point"]
}

3. Heartbeat Monitoring (Every 30 Minutes)

Automatic Background Process:

Every 30 minutes of conversation:

1. Check temporary memory buffer
2. Re-evaluate importance scores
   - Has this been mentioned again? (+0.15 boost)
   - Is it still relevant? (decay old items)
   - Any duplicates? (merge them)
3. Update statistics
4. Silently consolidate (no user notification)

Example:

[30 minutes pass]
AI Internal: "User mentioned 'prefers TypeScript' 3 times. Boosting importance from 0.85 to 0.95"
[Silent update - no notification]

4. Nightly Review (9 PM)

Active Engagement Process:

When current time >= 9 PM (or user says "review" or "复盘"):

1. Gather today's memories
2. Sort by importance (high → low)
3. Group by category
4. Present to user for confirmation

Format:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🌙 NIGHTLY MEMORY REVIEW
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Today I captured X memories:

🔴 HIGH IMPORTANCE (>0.85)
  ✓ "Use React, not Vue" (0.95)
  ✓ "Prefers dark mode" (0.90)

🟡 MEDIUM IMPORTANCE (0.60-0.85)
  ⚠ "API endpoint: /api/users" (0.70)
  ⚠ "Working on client project" (0.75)

🟢 LOW IMPORTANCE (<0.60)
  ? "Had pizza for lunch" (0.30)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Actions:
- Keep all high importance
- Review medium importance
- Discard low importance (unless you object)

Your call: Keep all? / Review specific? / Change anything?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

User Responses:

5. Memory Retrieval (Automatic)

Before responding to user:

1. Check: Is user asking about something I should remember?
2. Search memories by:
   - Keyword matching
   - Semantic similarity
   - Recency + importance score
3. If found:
   - Reference the memory: "Based on what you mentioned earlier..."
   - Use the information to inform response
4. If not found:
   - Answer normally (no mention of memory system)

Example:

User: "What framework should we use?"

AI Internal: [Searches memory → finds "Use React, not Vue" (0.95 importance)]

AI Response: "Based on your earlier decision, we should use React for this project."

Memory Categories

1. Preferences (0.90-1.0)

User likes/dislikes, habits, preferred ways of working.

Examples:

• "Prefers dark mode in IDEs"
• "Likes concise code comments"
• "Dislikes verbose logging"

2. Decisions (0.95-1.0)

Choices made, technical decisions, project directions.

Examples:

• "Use React, not Vue"
• "Deploy on Tuesdays only"
• "Database: PostgreSQL not MongoDB"

3. Goals (0.85-0.95)

Objectives, deadlines, targets.

Examples:

• "Launch beta by March 15"
• "Reduce API latency to <100ms"
• "Learn Rust this quarter"

4. Corrections (0.90-0.98)

User corrections, feedback, "I was wrong" moments.

Examples:

• "No, the API is /api/v2/users"
• "Actually, I meant TypeScript, not JavaScript"
• "That approach won't work because..."

5. Facts (0.60-0.80)

Technical details, context, information.

Examples:

• "API endpoint: /api/users"
• "Team size: 5 developers"
• "Client: Acme Corp"

6. Context (0.40-0.70)

Background information, situational context.

Examples:

• "Working on client project for Acme"
• "Preparing for demo next week"
• "Office is closed on Mondays"

Memory Storage Structure

storage/
├── conversations/           # Temporary conversation buffer
│   ├── 2026-03-19-morning.json
│   └── 2026-03-19-afternoon.json
│
├── memories/               # Permanent memory store
│   ├── preferences.json
│   ├── decisions.json
│   ├── goals.json
│   ├── corrections.json
│   ├── facts.json
│   └── context.json
│
└── reviews/                # Nightly review records
    ├── 2026-03-19-review.json
    └── 2026-03-18-review.json

Memory Format

{
  "id": "20260319-143052-abc123",
  "content": "User prefers dark mode",
  "category": "preference",
  "importance": 0.90,
  "source": "conversation",
  "timestamp": "2026-03-19T14:30:52Z",
  "context": "Mentioned during IDE setup",
  "mentioned_count": 3,
  "last_mentioned": "2026-03-19T16:45:00Z",
  "reviewed": false,
  "tags": ["ui", "preferences", "ide"]
}

Integration with Other Skills

Works Alongside:

• ✅ simple-memory-skill (complementary)
• ✅ Any coding skill (enhances context awareness)
• ✅ Any writing skill (remembers preferences)
• ✅ Any analysis skill (uses past context)

Non-Intrusive Design:

• Runs in background
• Silent unless important
• No extra setup required
• Works with any conversation type

Advanced Features

1. Importance Decay

Old memories gradually lose importance unless reinforced:

• 7 days old: ×0.9 multiplier
• 30 days old: ×0.7 multiplier
• 90 days old: ×0.5 multiplier

2. Repetition Boost

Each additional mention increases importance:

• 1st mention: base importance
• 2nd mention: +0.15
• 3rd mention: +0.10
• 4th+ mention: +0.05

3. Memory Relationships

Link related memories:

{
  "content": "Use React for frontend",
  "related_to": ["prefers-typescript", "component-architecture"]
}

4. Confidence Scores

Track certainty of memory:

{
  "confidence": 0.95,
  "source": "explicit_user_statement",
  "verified_count": 3
}

5. Temporary Memories

Mark time-sensitive info:

{
  "expires_at": "2026-03-25T00:00:00Z",
  "auto_archive": true
}

User Interactions

User Can Say:

AI Should Never:

• ❌ Annoy user with trivial memories
• ❌ Store sensitive information without warning
• ❌ Make assumptions about user intent
• ❌ Override explicit user corrections
• ❌ Store information user asked to forget

Configuration

User can customize (optional):

{
  "review_time": "21:00",           // Nightly review time
  "heartbeat_interval": 30,         // Minutes between heartbeats
  "importance_threshold": 0.70,     // Minimum importance to keep
  "auto_archive_days": 7,           // Days before auto-archive
  "enable_silent_mode": false,      // If true, suppress all notifications
  "categories": {
    "preferences": { "enabled": true, "min_importance": 0.90 },
    "decisions": { "enabled": true, "min_importance": 0.95 },
    "goals": { "enabled": true, "min_importance": 0.85 },
    "corrections": { "enabled": true, "min_importance": 0.90 }
  }
}

Troubleshooting

AI not remembering important things: → Check importance scoring logic → Verify storage is working → Review conversation logs

Too many trivial memories: → Adjust importance thresholds → Enable silent mode for low-importance items → Review nightly and discard

Nightly review not triggering: → Check current time vs configured review_time → Verify conversation activity today → Manual trigger: user says "review" or "复盘"

Memory retrieval fails: → Check search query matching → Verify memory file format → Review importance decay settings

Best Practices

For AI:

1. Be subtle - Don't mention memory system unless necessary
2. Be accurate - Verify information before storing
3. Be respectful - Don't store sensitive info without context
4. Be proactive - Anticipate what will be important later
5. Be concise - Store the essence, not the entire conversation

For Users:

1. Be explicit when something is really important
2. Correct the AI if it misremembers
3. Review nightly to curate your memory
4. Trust but verify - AI can make mistakes
5. Use keywords like "remember", "important", "critical" to boost detection

Performance Characteristics

• Storage: <1MB for 1000 memories
• Search time: <50ms for 10,000 memories
• Memory overhead: Minimal, runs in background
• Battery impact: Negligible (heartbeat only every 30min)

Comparison: Traditional vs Dynamic

Future Enhancements (Planned)

• Multi-user memory separation
• Memory encryption for sensitive data
• Cloud sync (optional)
• Memory export/import
• Visual memory graph
• Memory search by time range
• Natural language memory queries
• Memory sharing between sessions
• Integration with calendar/reminders
• Emotion detection in memories

License

MIT License - Free to use, modify, and distribute

Summary

Dynamic Memory System is a proactive, intelligent memory layer that:

✅ Automatically captures what matters ✅ Intelligently evaluates importance ✅ Proactively reviews daily memories ✅ Silently enhances any conversation ✅ Zero friction for the user

Result: You'll never have to say "remember this" again. The AI just *knows*.

Perfect for:

• Power users who want AI to remember everything
• Long-term projects spanning multiple sessions
• Complex work requiring consistent context
• Anyone tired of repeating themselves

Not ideal for:

• Single-use conversations (no overhead anyway)
• Users who want complete control (use manual memory instead)
• Highly sensitive work (use encrypted memory system)