mcp-server-memory by radek-paclt - MCP Server

MCP Memory Server

A human-like memory system using Qdrant vector database and OpenAI embeddings, accessible through the Model Context Protocol (MCP).

🔄 Automatické aktualizace

Rychlá aktualizace

./update.sh
# nebo
npm run update

Automatické sledování aktualizací

./watch-updates.sh
# nebo
npm run watch-updates

Skript kontroluje nové verze každých 5 minut a upozorní vás.

Features

Human-like Memory Types:
- Episodic (personal experiences)
- Semantic (facts and knowledge)
- Procedural (how to do things)
- Emotional (emotional memories)
- Sensory (sensory impressions)
- Working (short-term memory)
Memory Characteristics:
- Importance scoring (0-1)
- Emotional valence (-1 to 1)
- Associations between memories
- Context (location, people, mood, activity)
- Decay factor and access tracking
🎯 Customer Support Scoping:
- Company Level: Global knowledge accessible to all agents
- Customer Level: Customer-specific information across all interactions
- Interaction Level: Conversation-specific notes and context

Setup

Install dependencies:
```
npm install
```

Configure environment:

cp .env.example .env
# Edit .env with your settings

Start Qdrant (if using local):

docker run -p 6333:6333 -p 6334:6334 \
  --name qdrant-memory \
  -v $(pwd)/qdrant_storage:/qdrant/storage:z \
  qdrant/qdrant

Build the project:
```
npm run build
```

Testing with Claude Code

Use the MCP configuration file with Claude Code CLI:

# Basic usage
claude -p "Store a memory about today's meeting" --mcp-config claude-code-mcp.json

# Skip permissions for automation
claude -p "Search my memories" --mcp-config claude-code-mcp.json --dangerously-skip-permissions

# List available tools
claude -p "List available memory tools" --mcp-config claude-code-mcp.json

Available MCP tools (prefixed with mcp__memory__):

store_memory - Store a new memory
search_memories - Search for memories using natural language
get_memory - Retrieve a specific memory
update_memory - Update an existing memory
delete_memory - Delete a memory
analyze_memories - Analyze memory patterns

Customer Support Scoping

The memory system supports three-level hierarchical scoping for customer support teams:

Company Level (no customer_id, no interaction_id)

# Store global knowledge accessible to all agents
claude -p 'Store memory: "New refund policy: Premium customers get instant refunds, standard customers within 3 business days"' --mcp-config claude-code-mcp.json

Customer Level (customer_id only)

# Store customer-specific information
claude -p 'Store memory: "Customer Sarah Johnson prefers phone support, has enterprise account, key contact for TechCorp implementation" with context: {"customer_id": "cust_12345", "tags": ["vip", "enterprise"]}' --mcp-config claude-code-mcp.json

Interaction Level (both customer_id and interaction_id)

# Store conversation-specific notes
claude -p 'Store memory: "Customer reports login issues with 2FA, helped reset authenticator app, issue resolved" with context: {"customer_id": "cust_12345", "interaction_id": "call_789", "tags": ["support", "resolved"]}' --mcp-config claude-code-mcp.json

Smart Search with Scope Inheritance

# Search inherits from all relevant scopes automatically
claude -p 'Search memories for "login issues" with customer_id: "cust_12345", interaction_id: "call_789"' --mcp-config claude-code-mcp.json

# Results include:
# 1. Current interaction memories (highest priority)
# 2. Customer-specific memories (medium priority)  
# 3. Company knowledge base (lowest priority, but still relevant)

Customer Support Use Cases

# Agent preparation before call
claude -p 'Search customer history for customer_id: "cust_12345"' --mcp-config claude-code-mcp.json

# During live support - get full context
claude -p 'Search memories for "billing" with customer_id: "cust_12345", interaction_id: "call_current"' --mcp-config claude-code-mcp.json

# Store resolution for team learning
claude -p 'Store memory: "Billing API timeout fixed by increasing database connection pool from 10 to 50 connections"' --mcp-config claude-code-mcp.json

Benefits

🔒 Automatic Isolation: Customer data stays separate
📈 Contextual Relevance: Right information at the right time
🧠 Team Learning: Company knowledge grows with each resolution
⚡ Fast Context: Instant access to customer history and current interaction
🔄 Inheritance: Broader knowledge automatically included when relevant

Migration from Previous Version

All existing memories automatically become "company level" - no breaking changes! Just start adding customer_id and interaction_id to new memories for scoping.

Example Usage

// Store a company-level memory (available to all agents)
store_memory({
  "content": "New API rate limits: 1000 requests per minute for free tier, 10000 for premium",
  "type": "semantic",
  "context": {
    "tags": ["api", "rate-limits", "policy"]
  },
  "importance": 0.9
})

// Store customer-specific memory
store_memory({
  "content": "Customer prefers email communication, works in EST timezone",
  "type": "semantic", 
  "context": {
    "customer_id": "cust_12345",
    "tags": ["preferences", "timezone"]
  },
  "importance": 0.7
})

// Store interaction-specific memory
store_memory({
  "content": "Resolved billing discrepancy of $45.32, issued refund",
  "type": "episodic",
  "context": {
    "customer_id": "cust_12345",
    "interaction_id": "call_789",
    "tags": ["billing", "refund", "resolved"]
  },
  "importance": 0.8
})

// Search with scoping
search_memories({
  "query": "billing issues",
  "customer_id": "cust_12345",
  "interaction_id": "call_789",
  "limit": 10
})

Development

npm run dev - Run in development mode
npm run build - Build TypeScript
npm run test - Run tests
npm run lint - Run linter
npm run typecheck - Type check