dive-memory-v3（潜水内存 v3）

Dive-Memory v3: MCP-Based Persistent Memory System

Dive-Memory v3 provides long-term persistent memory for AI agents, solving the "context forgetting" problem across sessions.

Core Capabilities

1. Memory Storage

Store memories with rich metadata using the Python API:

from dive_memory_v3 import DiveMemory

memory = DiveMemory()

# Add memory
memory.add(
    content="Fixed JWT auth bug with refresh token rotation",
    section="solutions",
    subsection="authentication",
    tags=["jwt", "security", "bug-fix"],
    importance=8,
    metadata={"code_snippet": "...", "success_rate": 1.0}
)

2. Semantic Search

Search using natural language + hybrid search (vector + keyword):

# Search memories
results = memory.search(
    query="How to fix JWT authentication issues?",
    section="solutions",
    tags=["authentication"],
    top_k=5
)

for result in results:
    print(f"[{result.importance}] {result.content}")
    print(f"Relevance: {result.score:.2f}")

3. Knowledge Graph

Automatically build relationships between memories:

# Get related memories
related = memory.get_related(memory_id, max_depth=2)

# Visualize graph
graph = memory.get_graph(section="solutions")
# Returns: {nodes: [...], edges: [...]}

4. Context Injection

Automatically inject relevant memories into prompts:

# Enable auto-injection
memory.enable_context_injection()

# When processing task, relevant memories auto-prepend
task = "Implement user authentication"
context = memory.get_context_for_task(task)
# Returns: "Past solutions: JWT with refresh tokens..."

5. Deduplication

Automatically detect and merge duplicate memories:

# Run deduplication
duplicates = memory.find_duplicates(threshold=0.95)
memory.merge_duplicates(duplicates, strategy="keep_newer")

6. Cloud Sync

Sync memories across devices:

# Configure cloud sync
memory.configure_sync(
    provider="s3",
    bucket="dive-memory-sync",
    auto_sync=True
)

# Manual sync
memory.sync_to_cloud()
memory.sync_from_cloud()

MCP Server Integration

Dive-Memory v3 runs as an MCP server for integration with Claude Desktop, Claude Code, etc.

Start MCP Server

cd /home/ubuntu/skills/dive-memory-v3/scripts
python3 mcp_server.py

MCP Tools Available

• memory_add: Add new memory
• memory_search: Search memories
• memory_update: Update existing memory
• memory_delete: Delete memory
• memory_graph: Get knowledge graph
• memory_related: Find related memories
• memory_stats: Get memory statistics

MCP Configuration

Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "dive-memory": {
      "command": "python3",
      "args": ["/home/ubuntu/skills/dive-memory-v3/scripts/mcp_server.py"],
      "env": {
        "OPENAI_API_KEY": "your-key-here"
      }
    }
  }
}

Memory Organization

Sections & Subsections

Organize memories hierarchically:

solutions/
  ├── authentication/
  ├── database/
  └── api/
decisions/
  ├── architecture/
  └── technology/
preferences/
research/
  ├── ai-models/
  └── frameworks/

Metadata Fields

• tags: List of keywords
• importance: 1-10 score
• source: Origin of memory
• timestamp: Creation time
• access_count: Usage frequency
• last_accessed: Last retrieval time

Use Cases

1. Coding Agent

Remember successful solutions and patterns:

# Store solution
memory.add(
    content="Use tRPC for type-safe APIs without code generation",
    section="solutions/api",
    tags=["typescript", "api", "type-safety"],
    importance=9
)

# Later, when building API
context = memory.search("How to build type-safe API?")

2. Research Agent

Build knowledge base from research:

# Store findings
memory.add(
    content="Claude Opus 4.5: Best for code quality (10/10)",
    section="research/ai-models",
    tags=["claude", "code-review"],
    importance=8
)

# Auto-link to related memories
# Links to: "GPT-5.2 for security", "DeepSeek for reasoning"

3. Decision Tracking

Remember architectural decisions:

memory.add(
    content="Chose PostgreSQL over MongoDB for ACID guarantees",
    section="decisions/database",
    tags=["database", "architecture"],
    metadata={"rationale": "Need transactions for financial data"}
)

4. Learning Loop

Learn from task execution:

# After successful task
memory.add(
    content="Agent #42 excels at React component refactoring",
    section="capabilities",
    tags=["agent-42", "react", "refactoring"],
    importance=7
)

# Route future React tasks to Agent #42

Advanced Features

Importance Scoring

Automatic importance calculation based on:

• Access frequency
• Recency
• Graph centrality (how connected)
• User-defined importance

Memory Pruning

Remove low-value memories:

# Prune memories with:
# - importance < 3
# - not accessed in 90 days
# - access_count < 2
memory.prune(
    min_importance=3,
    max_age_days=90,
    min_access_count=2
)

Memory Consolidation

Merge similar memories:

# Find similar memories (0.7-0.95 similarity)
similar = memory.find_similar(threshold=0.7)

# Consolidate into summary
memory.consolidate(similar, strategy="llm_summary")

Export & Import

# Export to JSON
memory.export("memories.json", section="solutions")

# Import from JSON
memory.import_from_json("memories.json")

# Export to Markdown
memory.export_markdown("knowledge_base.md")

Performance

• Search: < 100ms for 10K memories
• Storage: Supports 1M+ memories
• Deduplication: < 1% false positives
• Cloud Sync: Background, non-blocking

Configuration

Configuration file at references/config.json contains all settings. Key options:

• Storage backend (SQLite/PostgreSQL)
• Embedding provider (OpenAI/local)
• Search strategy (semantic/keyword/hybrid)
• Deduplication thresholds
• Cloud sync settings

See references/config.json for full configuration options.

Scripts Reference

• scripts/mcp_server.py: MCP server entry point
• scripts/memory_cli.py: Command-line interface
• scripts/setup_database.py: Initialize SQLite database
• scripts/sync_to_cloud.py: Manual cloud sync
• scripts/export_graph.py: Export knowledge graph visualization

Best Practices

1. Use sections: Organize memories by domain
2. Tag generously: Enable better search
3. Set importance: Help prioritize retrieval
4. Enable auto-injection: Reduce manual context management
5. Regular deduplication: Keep memory clean
6. Cloud sync: Backup and multi-device access
7. Prune old memories: Prevent database bloat

Troubleshooting

Search returns no results

• Check section/tag filters
• Try broader query
• Verify embeddings are generated

Slow search performance

• Run VACUUM on SQLite database
• Reduce top_k parameter
• Enable query caching

Duplicate memories not detected

• Lower similarity_threshold
• Check embedding quality
• Verify content normalization

Integration with Dive AI

Dive-Memory v3 integrates seamlessly with Dive AI V20:

from dive_ai import DiveOrchestrator
from dive_memory_v3 import DiveMemory

# Initialize
orchestrator = DiveOrchestrator()
memory = DiveMemory()

# Enable memory for orchestrator
orchestrator.set_memory(memory)

# Execute task with auto-context injection
result = orchestrator.execute(
    task="Build authentication system",
    use_memory=True  # Auto-inject relevant memories
)

# Store execution results
memory.add(
    content=f"Task completed: {result.summary}",
    section="executions",
    tags=["authentication", "success"],
    metadata={"cost": result.cost, "time": result.duration}
)

References

• Full API documentation: references/api_reference.md
• MCP protocol spec: references/mcp_protocol.md
• Database schema: references/schema.sql
• Configuration guide: references/config.json