Memory Design Patterns

Production-ready memory architecture patterns for AI applications using Mem0. This skill provides comprehensive guidance on designing scalable, performant memory systems with proper isolation, retention strategies, and optimization techniques.

Instructions

Phase 1: Understand Memory Types

Mem0 provides three distinct memory scopes, each serving different purposes:

1. User Memory (Persistent Preferences & Profile)

Purpose: Long-term personal preferences, profile data, and user characteristics that persist across all interactions.

Use Cases:

• User preferences (dietary restrictions, communication style, language preferences)
• Personal information (location, occupation, family details)
• Long-term goals and interests
• Historical context that should persist indefinitely

Implementation:

# Add user-level memory
memory.add(
    "User prefers concise responses without technical jargon"
    user_id="customer_bob"
)

# Search user memories
user_context = memory.search(
    "communication style"
    user_id="customer_bob"
)

Key Characteristics:

• Persists indefinitely (or until explicitly deleted)
• Shared across all agents interacting with this user
• Should contain stable, long-term information
• Typically 10-50 memories per user

2. Agent Memory (Agent-Specific Context)

Purpose: Agent-specific knowledge, behaviors, and learned patterns that apply across all users interacting with this agent.

Use Cases:

• Agent capabilities and limitations
• Domain-specific knowledge
• Learned behaviors and patterns
• Agent-specific instructions and protocols

Implementation:

# Add agent-level memory
memory.add(
    "When handling refund requests, always check order date first"
    agent_id="support_agent_v2"
)

# Search agent memories
agent_context = memory.search(
    "refund process"
    agent_id="support_agent_v2"
)

Key Characteristics:

• Shared across all users interacting with this agent
• Contains agent-specific procedures and knowledge
• Moderate retention (days to months)
• Typically 50-200 memories per agent

3. Session/Run Memory (Temporary Conversation Context)

Purpose: Ephemeral context specific to a single conversation or task session.

Use Cases:

• Current conversation topic
• Temporary task context
• Session-specific state
• Short-term working memory

Implementation:

# Add session-level memory
memory.add(
    "Current issue: payment failed with error code 402"
    run_id="session_12345_20250115"
)

# Search session memories
session_context = memory.search(
    "current issue"
    run_id="session_12345_20250115"
)

Key Characteristics:

• Short-lived (minutes to hours)
• Isolated to specific conversation or task
• Should be cleaned up after session ends
• Typically 5-20 memories per session

Phase 2: Choose Storage Backend (Vector vs Graph)

Vector Memory (Default)

How It Works: Embeddings stored in vector database, semantic similarity search using cosine distance.

Strengths:

• Fast semantic search
• Excellent for unstructured data
• Low setup complexity
• Works out-of-the-box with Mem0

Weaknesses:

• Cannot query relationships
• No explicit entity connections
• Limited reasoning about connections

Best For:

• Simple preference storage
• Document/chunk retrieval
• Semantic search use cases
• Quick prototyping

Configuration:

from mem0 import Memory

# Default vector-only configuration
memory = Memory()

Graph Memory (Advanced)

How It Works: Entities and relationships stored in graph database (Neo4j/Memgraph), enables relationship traversal and complex queries.

Strengths:

• Explicit entity relationships
• Complex query capabilities
• Relationship reasoning
• Multi-hop traversal

Weaknesses:

• Requires graph database setup
• Higher infrastructure complexity
• Slower for pure semantic search
• More storage overhead

Best For:

• Multi-entity systems
• Relationship-heavy domains
• Complex reasoning requirements
• Enterprise knowledge graphs

Configuration:

from mem0 import Memory
from mem0.configs.base import MemoryConfig

config = MemoryConfig(
    graph_store={
        "provider": "neo4j"
        "config": {
            "url": "bolt://localhost:7687"
            "username": "neo4j"
            "password": "password"
        }
    }
)
memory = Memory(config)

Decision Matrix:

Use Case	Vector	Graph
User preferences	✅ Best	⚠️ Overkill
Product recommendations	✅ Best	⚠️ Overkill
Customer support	✅ Good	✅ Better
Knowledge management	⚠️ Limited	✅ Best
Multi-tenant systems	✅ Good	✅ Best
Team collaboration	⚠️ Limited	✅ Best

Phase 3: Design Retention Strategy

Use the retention strategy template:

bash scripts/generate-retention-policy.sh <memory-type> <retention-days>

Retention Guidelines

User Memory:

• Retention: Indefinite (with user control)
• Cleanup: User-initiated deletion only
• Archival: After 1 year of inactivity
• GDPR: Must support right to deletion

Agent Memory:

• Retention: 90-180 days typical
• Cleanup: Automatic based on relevance score
• Versioning: Keep agent version history
• Deprecation: Clear old agent memories on major updates

Session Memory:

• Retention: 1-24 hours
• Cleanup: Automatic after session end
• Conversion: Promote important memories to user/agent level
• Storage: Consider in-memory for very short sessions

Retention Implementation

Run the retention analyzer:

bash scripts/analyze-retention.sh <user_id_or_agent_id>

This script:

1. Analyzes memory age and access patterns
2. Identifies stale memories
3. Suggests cleanup actions
4. Generates retention reports

Phase 4: Implement Multi-Level Memory Pattern

Pattern: Combine all three memory types for comprehensive context.

Template: Use

templates/multi-level-memory-pattern.py

Architecture:

Query Processing Flow:
1. Retrieve session context (immediate)
2. Retrieve user context (preferences)
3. Retrieve agent context (capabilities)
4. Merge contexts with priority weighting
5. Generate response with full context

Priority Weighting:

• Session: 40% weight (most relevant to current task)
• User: 35% weight (personalizes response)
• Agent: 25% weight (ensures consistent behavior)

Implementation:

# Retrieve all context levels
session_memories = memory.search(query, run_id=run_id)
user_memories = memory.search(query, user_id=user_id)
agent_memories = memory.search(query, agent_id=agent_id)

# Weighted merge
context = merge_contexts(
    session=session_memories
    user=user_memories
    agent=agent_memories
    weights={"session": 0.4, "user": 0.35, "agent": 0.25}
)

Phase 5: Optimize Performance

Vector Search Optimization

Run the performance analyzer:

bash scripts/analyze-memory-performance.sh <project_name>

Optimization Techniques:

1. Limit Search Results:

   memories = memory.search(query, user_id=user_id, limit=5)
   

   • Default: 10 results
   • Recommended: 3-5 for chat, 10-20 for RAG

2. Use Filters to Reduce Search Space:

   memories = memory.search(
       query
       filters={
           "AND": [
               {"user_id": "alex"}
               {"agent_id": "support_agent"}
           ]
       }
   )
   

3. Cache Frequently Accessed Memories:

   • Cache user preferences (rarely change)
   • Refresh cache every 5-10 minutes
   • Invalidate on explicit memory updates

4. Batch Operations:

   # Add multiple memories in one call
   memory.add(messages, user_id=user_id)
   

Graph Query Optimization

For graph memory:

1. Limit Traversal Depth: Max 2-3 hops
2. Index Key Properties: user_id, agent_id, timestamps
3. Use Relationship Filters: Reduce unnecessary traversals
4. Monitor Query Performance: Track slow queries > 100ms

Phase 6: Implement Cost Optimization

Run the cost analyzer:

bash scripts/analyze-memory-costs.sh <user_id> <date_range>

Cost Optimization Strategies:

1. Deduplication: Remove similar/redundant memories

   bash scripts/deduplicate-memories.sh <user_id>
   

2. Archival: Move old memories to cold storage

   • Active: Last 30 days (vector DB)
   • Archive: 30-180 days (compressed JSON)
   • Long-term: > 180 days (S3/cold storage)

3. Compression: Use shorter embeddings for less critical memories

   • Critical: 1536 dimensions (OpenAI large)
   • Standard: 768 dimensions (OpenAI small)
   • Archival: 384 dimensions (lightweight model)

4. Smart Pruning: Remove low-value memories

   • Score-based: Keep only high relevance scores
   • Access-based: Remove never-accessed memories
   • Importance-based: User/agent priority tagging

Phase 7: Security and Isolation

Multi-Tenant Isolation

Pattern: Ensure complete data isolation between users/organizations.

Implementation:

# Always scope by user_id or org_id
memories = memory.search(
    query
    filters={"user_id": current_user_id}
)

# Validate access before retrieval
if not user_has_access(user_id, requested_user_id):
    raise PermissionError("Access denied")

Security Checklist:

• ✅ Never allow cross-user memory access
• ✅ Validate all user_id parameters
• ✅ Implement org-level isolation for multi-tenant apps
• ✅ Audit memory access logs
• ✅ Encrypt sensitive memory content
• ✅ Support GDPR right to deletion

Run the security audit:

bash scripts/audit-memory-security.sh

Decision Trees

When to Use Each Memory Type

Use the decision helper:

bash scripts/suggest-memory-type.sh "<use_case_description>"

Quick Reference:

• User dietary preferences → User Memory
• Agent's SOP for task X → Agent Memory
• Current conversation topic → Session Memory
• Customer support ticket details → Session Memory (promote to User if resolved)
• System capabilities → Agent Memory
• User's birthday → User Memory

Vector vs Graph Decision

Use the architecture advisor:

bash scripts/suggest-storage-architecture.sh "<project_description>"

Decision Criteria:

1. Need relationship traversal? → Graph
2. Pure semantic search? → Vector
3. < 10,000 memories total? → Vector
4. Complex entity relationships? → Graph
5. Team/org hierarchies? → Graph
6. Simple preference storage? → Vector

Key Files

Scripts (all functional, not placeholders):

• scripts/generate-retention-policy.sh

  - Create retention policy configs

• scripts/analyze-retention.sh

  - Analyze memory age and access patterns

• scripts/analyze-memory-performance.sh

  - Performance profiling

• scripts/analyze-memory-costs.sh

  - Cost analysis and optimization suggestions

• scripts/deduplicate-memories.sh

  - Find and remove duplicate memories

• scripts/audit-memory-security.sh

  - Security compliance checking

• scripts/suggest-memory-type.sh

  - Interactive memory type advisor

• scripts/suggest-storage-architecture.sh

  - Architecture recommendation tool

Templates:

• templates/multi-level-memory-pattern.py

  - Complete implementation

• templates/retention-policy.yaml

  - Retention configuration

• templates/vector-only-config.py

  - Vector memory setup

• templates/graph-memory-config.py

  - Graph memory setup

• templates/hybrid-architecture.py

  - Vector + Graph combined

• templates/cost-optimization-config.yaml

  - Cost optimization settings

Examples:

• examples/customer-support-memory-architecture.md

  - Full implementation guide

• examples/multi-agent-collaboration.md

  - Shared memory patterns

• examples/e-commerce-personalization.md

  - Product recommendation memory

• examples/healthcare-assistant.md

  - HIPAA-compliant memory architecture

Best Practices

1. Start Simple: Use vector-only with user + session memories
2. Add Complexity as Needed: Only introduce graph when relationships matter
3. Monitor Performance: Track memory retrieval times and costs
4. Implement Retention Early: Don't let memory grow unbounded
5. Test Isolation: Verify cross-user memory access is impossible
6. Document Memory Schema: Track what memories mean and when they're used
7. Version Agent Memories: Clear separation between agent versions
8. Promote Important Memories: Session → User when patterns emerge
9. Use Metadata: Tag memories with categories for better filtering
10. Regular Audits: Monthly review of memory growth and costs

Troubleshooting

Slow Memory Retrieval:

• Reduce search limit
• Add more specific filters
• Check vector index performance
• Consider caching

High Costs:

• Run cost analyzer script
• Implement deduplication
• Review retention policy
• Archive old memories

Poor Search Results:

• Check embedding model quality
• Verify memory content is descriptive
• Use hybrid search (keyword + semantic)
• Add metadata for filtering

Memory Leakage Between Users:

• Audit security script immediately
• Review all memory queries for user_id filtering
• Check RLS policies if using custom backends
• Implement access logging

---

Plugin: mem0 Version: 1.0.0 Last Updated: 2025-10-27