Custom Prompts Usage Guide

This guide provides detailed instructions on how to use custom prompts in powermem to customize the behavior of various memory operations.

Overview

powermem supports customizing prompts at multiple levels, including:

1. Fact Extraction - Extract factual information from conversations
2. Memory Update - Determine how to update existing memories
3. Importance Evaluation - Evaluate the importance of memories
4. Graph Memory Operations - Operations on graph-structured memories (extract relations, update, delete)

Configuration Methods

Method 1: Using MemoryConfig Object (Recommended)

from powermem import Memory, MemoryConfig, GraphStoreConfig, LlmConfig, EmbedderConfig, VectorStoreConfig, IntelligentMemoryConfig

# Create configuration object
config = MemoryConfig(
    # Custom prompt for fact extraction
    custom_fact_extraction_prompt="""
You are an information extraction expert. Extract user preferences, important facts, and plans from conversations.

Rules:
1. Only extract explicitly mentioned information
2. Preserve temporal information (e.g., "yesterday", "last month")
3. Extract related facts separately

Output JSON format: {"facts": ["fact1", "fact2"]}
""",
    
    # Custom prompt for memory update
    custom_update_memory_prompt="""
You are a memory manager. Compare new facts with existing memories and decide: ADD (new), UPDATE (update), DELETE (delete), or NONE (no change).

Important rules:
- If new fact contains temporal info that old memory doesn't, then UPDATE
- If information completely conflicts, then DELETE
- If information is the same, then NONE
""",
    
    # Custom prompt for importance evaluation
    custom_importance_evaluation_prompt="""
Evaluate the importance of memory content on a scale from 0.0 to 1.0.

Evaluation criteria:
- Relevance: How relevant to user needs
- Novelty: Information uniqueness
- Emotional Impact: Emotional significance
- Actionability: Information usefulness

Return JSON format:
{
    "importance_score": <0.0-1.0>,
    "reasoning": "<explanation>",
    "criteria_scores": {
        "relevance": <score>,
        "novelty": <score>,
        "emotional_impact": <score>,
        "actionable": <score>
    }
}
""",
    
    # Graph memory configuration (optional)
    graph_store=GraphStoreConfig(
        enabled=True,
        
        # Custom prompt for relation extraction
        custom_extract_relations_prompt="""
Extract entities and relationships from text to build a knowledge graph.

Special rules:
- Use "USER_ID" as the entity for user self-references (e.g., "I", "me", "my")
- Only extract explicitly mentioned relationships
- Maintain consistency and generality in relationship types

Extracted relationships should follow this format:
source -- relationship -- destination
""",
        
        # Custom prompt for graph update
        custom_update_graph_prompt="""
Analyze existing graph memories and new information, update relationships.

Existing memories:
{existing_memories}

New memories:
{new_memories}

Rules:
1. If new information conflicts with existing memory, update the relationship
2. If new information is more detailed, update the relationship
3. Maintain graph structure consistency
""",
        
        # Custom prompt for deleting relations
        custom_delete_relations_prompt="""
Analyze existing relationships and determine which should be deleted.

Existing relationships:
- Only delete outdated or conflicting relationships
- If new information contradicts existing relationship, then delete
- Do not delete if there is a possibility of same type relationship with different destination nodes

Example:
Existing: Alice -- loves_to_eat -- pizza
New Information: Alice also loves to eat burger
Result: Do not delete (Alice can love both pizza and burger)
""",
        
        # Or use generic prompt (for relation extraction)
        # custom_prompt="Your custom prompt..."
    ),
    
    # Other configurations...
    llm=LlmConfig(provider="openai", config={"api_key": "..."}),
    embedder=EmbedderConfig(provider="openai", config={"api_key": "..."}),
    vector_store=VectorStoreConfig(provider="chroma", config={})
)

# Create Memory instance
memory = Memory(config=config)

Method 2: Using Dictionary Configuration

from powermem import Memory

# Dictionary configuration (no need to import config classes)

config = {
    # Custom prompt for fact extraction
    "custom_fact_extraction_prompt": """
You are an information extraction expert. Extract user preferences, important facts, and plans from conversations.
Output JSON format: {"facts": ["fact1", "fact2"]}
""",
    
    # Custom prompt for memory update
    "custom_update_memory_prompt": """
You are a memory manager. Compare new facts with existing memories and decide: ADD, UPDATE, DELETE, or NONE.
""",
    
    # Custom prompt for importance evaluation
    "custom_importance_evaluation_prompt": """
Evaluate the importance of memory content on a scale from 0.0 to 1.0.
Return JSON format containing importance_score and reasoning.
""",
    
    # Graph memory configuration
    "graph_store": {
        "enabled": True,
        "custom_extract_relations_prompt": "Extract entities and relationships from text...",
        "custom_update_graph_prompt": "Update graph memories...",
        "custom_delete_relations_prompt": "Delete relationships...",
        # Or
        "custom_prompt": "Generic prompt (for relation extraction)"
    },
    
    # Other configurations...
    "llm": {
        "provider": "openai",
        "config": {"api_key": "..."}
    },
    "embedder": {
        "provider": "openai",
        "config": {"api_key": "..."}
    },
    "vector_store": {
        "provider": "chroma",
        "config": {}
    }
}

memory = Memory(config=config)

Detailed Examples

Example 1: Custom Fact Extraction

from powermem import Memory, MemoryConfig, LlmConfig, EmbedderConfig, VectorStoreConfig

# Create a prompt focused on extracting preferences and interests
config = MemoryConfig(
    custom_fact_extraction_prompt="""
You are a preferences and interests extraction expert.

Extract from conversations:
1. User preferences (things they like/dislike)
2. Hobbies and interests
3. Lifestyle choices

Format requirements:
- Each fact should be independent and complete sentence
- Preserve specific details (brands, locations, times)
- Output JSON: {"facts": ["fact list"]}

Conversation content:
""",
    llm=LlmConfig(provider="openai", config={"api_key": "..."}),
    embedder=EmbedderConfig(provider="openai", config={"api_key": "..."}),
    vector_store=VectorStoreConfig(provider="chroma", config={})
)

memory = Memory(config=config)

# Add memory, system will use custom fact extraction prompt
result = memory.add(
    messages=[
        {"role": "user", "content": "I like latte coffee, I need a cup every morning."},
        {"role": "assistant", "content": "Got it, I've remembered your coffee preference."}
    ],
    user_id="user123",
    infer=True  # Enable intelligent extraction
)

print(result)

Example 2: Custom Memory Update Strategy

from powermem import Memory, MemoryConfig, LlmConfig, EmbedderConfig, VectorStoreConfig

# Create a more conservative update strategy
config = MemoryConfig(
    custom_update_memory_prompt="""
You are a memory manager with a conservative update strategy.

Operation rules:
1. ADD: New information → Add
2. UPDATE: Only update if new information is more specific and detailed
3. DELETE: Only delete if clearly conflicting
4. NONE: Information is same or similar → Keep unchanged

Temporal information handling:
- If new fact has date but old memory doesn't → UPDATE (add time)
- If dates conflict → Keep the latest

Memory format:
{"id": "<id>", "text": "<content>", "event": "ADD|UPDATE|DELETE|NONE"}

Existing memories:
{old_memory}

New facts:
{new_facts}
""",
    llm=LlmConfig(provider="openai", config={"api_key": "..."}),
    embedder=EmbedderConfig(provider="openai", config={"api_key": "..."}),
    vector_store=VectorStoreConfig(provider="chroma", config={})
)

memory = Memory(config=config)

# System will use custom update strategy
memory.add(messages="I like Python programming", user_id="user123")
memory.add(messages="I love Python and JavaScript programming", user_id="user123")  # Will update instead of add new

Example 3: Custom Importance Evaluation

from powermem import Memory, MemoryConfig, IntelligentMemoryConfig, LlmConfig, EmbedderConfig, VectorStoreConfig

# Create an evaluation prompt that emphasizes emotional value and practicality
config = MemoryConfig(
    custom_importance_evaluation_prompt="""
Evaluate the importance of memory content, focusing on emotional value and practicality.

Scoring weights:
- Emotional Impact: 40% (highest weight)
- Actionability: 30%
- Relevance: 20%
- Novelty: 10%

Scoring criteria:
- 0.8-1.0: Extremely high importance (e.g., passwords, important dates, health information)
- 0.6-0.8: High importance (e.g., preferences, plans, relationships)
- 0.4-0.6: Medium importance (e.g., general information)
- 0.0-0.4: Low importance (e.g., temporary information)

Content: {content}
Metadata: {metadata}

Return JSON:
{{
    "importance_score": <0.0-1.0>,
    "reasoning": "<detailed explanation>",
    "criteria_scores": {{
        "emotional_impact": <0.0-1.0>,
        "actionable": <0.0-1.0>,
        "relevance": <0.0-1.0>,
        "novelty": <0.0-1.0>
    }}
}}
""",
    intelligent_memory=IntelligentMemoryConfig(
        enabled=True
    ),
    llm=LlmConfig(provider="openai", config={"api_key": "..."}),
    embedder=EmbedderConfig(provider="openai", config={"api_key": "..."}),
    vector_store=VectorStoreConfig(provider="chroma", config={})
)

memory = Memory(config=config)

# System will use custom importance evaluation
memory.add(
    messages="I have diabetes and need to monitor blood sugar control",
    user_id="user123",
    metadata={"category": "health", "priority": "critical"}
)

Example 4: Custom Graph Memory Operations

from powermem import Memory, MemoryConfig, GraphStoreConfig, LlmConfig, EmbedderConfig, VectorStoreConfig

config = MemoryConfig(
    graph_store=GraphStoreConfig(
        enabled=True,
        provider="oceanbase",
        
        # Relation extraction: Focus on extracting person relationships
        custom_extract_relations_prompt="""
Extract entities and relationships from text, with special focus on person relationships.

Extraction rules:
1. Identify all person entities
2. Extract relationships between persons (e.g., friend, colleague, family member)
3. Extract relationships between persons and locations, events
4. Use "USER_ID" to represent the user themselves ("I", "my")

Relationship type examples:
- Interpersonal: friend_of, colleague_of, family_member_of
- Location: lives_in, works_at, visited
- Event: attended, organized, participated_in

Text: {text}
""",
        
        # Graph update: Intelligent relationship merging
        custom_update_graph_prompt="""
Update knowledge graph, intelligently merge and update relationships.

Existing relationships:
{existing_memories}

New relationships:
{new_memories}

Update strategy:
1. If relationship is same but destination is different, keep both (e.g., Alice -- likes -- pizza and Alice -- likes -- burger)
2. If relationship is more specific, update relationship description
3. If time is more recent, update to latest information
4. Avoid duplicates and redundancy
""",
        
        # Delete relationships: Conservative strategy
        custom_delete_relations_prompt="""
Determine relationships to delete using conservative strategy.

Deletion conditions (all must be met):
1. New information explicitly negates existing relationship
2. New information is more recent and accurate
3. Relationship type and destination are both the same

Do not delete if:
- Relationship type is same but destination is different (allow multiple destinations)
- Information is ambiguous or unclear
- Could be supplementary information rather than conflict

Existing relationship list:
- Each relationship format: source -- relationship -- destination

New information:
{new_text}

Only return relationships that need to be deleted.
""",
        
        config={
            "host": "127.0.0.1",
            "port": 2881,
            "user": "root",
            "password": "password",
            "db_name": "test",
            "embedding_model_dims": 768
        }
    ),
    llm=LlmConfig(provider="openai", config={"api_key": "..."}),
    embedder=EmbedderConfig(provider="openai", config={"api_key": "..."}),
    vector_store=VectorStoreConfig(provider="chroma", config={})
)

memory = Memory(config=config)

# Add memory, system will use custom graph memory prompts
memory.add(
    messages="Alice is my friend, we work together at Google. She likes to drink coffee.",
    user_id="user123",
    agent_id="agent1"
)

Prompt Placeholders

Different prompts support different placeholders:

Fact Extraction Prompt

• No specific placeholders needed, system will automatically add conversation content

Memory Update Prompt

• {old_memory} or get existing memories from config
• {new_facts} or get new facts from config

Importance Evaluation Prompt

• {content} - Content to evaluate
• {metadata} - Metadata (optional)
• {context} - Context (optional)

Graph Update Prompt

• {existing_memories} - Existing graph memories
• {new_memories} - New graph memories

Relation Extraction Prompt

• USER_ID - Will be replaced with actual user ID
• CUSTOM_PROMPT - If custom_prompt is provided, this placeholder will be replaced

Delete Relations Prompt

• USER_ID - Will be replaced with actual user ID

Prompt Writing Guidelines

1. Clarity

• Use clear, specific instructions
• Avoid ambiguous expressions

2. Format Specification

• Clearly specify output format (JSON, etc.)
• Provide example formats

3. Edge Cases

• Explain how to handle conflicting information
• Define conditions for deletion/update

4. Context Preservation

• Require preservation of temporal information
• Require preservation of relevant details

5. Consistency

• Use consistent terminology
• Maintain uniform format

Complete Example: Multi-language Support

from powermem import Memory, MemoryConfig, LlmConfig, EmbedderConfig, VectorStoreConfig

# Chinese prompt configuration
chinese_config = MemoryConfig(
    custom_fact_extraction_prompt="""
You are a Chinese information extraction expert. Extract user information from Chinese conversations.

Extraction rules:
1. Extract personal information, preferences, plans
2. Preserve temporal expressions (e.g., "yesterday", "last week")
3. Maintain naturalness of Chinese expressions

Output format: {"facts": ["fact1", "fact2"]}
""",
    
    custom_update_memory_prompt="""
You are a Chinese memory manager. Compare new facts with existing memories.

Operations: ADD (新增), UPDATE (更新), DELETE (删除), NONE (不变)

Important: Preserve natural Chinese expressions.
""",
    llm=LlmConfig(provider="qwen", config={"api_key": "..."}),
    embedder=EmbedderConfig(provider="qwen", config={"api_key": "..."}),
    vector_store=VectorStoreConfig(provider="oceanbase", config={})
)

memory_cn = Memory(config=chinese_config)

# Use Chinese conversations
memory_cn.add(
    messages=[
        {"role": "user", "content": "我昨天去了北京，参观了故宫。"},
        {"role": "assistant", "content": "听起来很有意思！"}
    ],
    user_id="user_cn",
    infer=True
)

Best Practices

1. Test Prompts: Test prompt effectiveness with small datasets before production use
2. Iterative Optimization: Continuously optimize prompts based on results
3. Version Control: Save prompts in configuration files for version management
4. Documentation: Document the purpose and special rules of each custom prompt
5. Performance Considerations: Concise prompts can reduce token consumption and improve speed

Important Notes

1. Backward Compatibility: If custom prompts are not provided, system will use default prompts
2. Format Requirements: Ensure custom prompts contain necessary placeholders
3. LLM Compatibility: Some prompt formats may be more effective for specific LLMs
4. Configuration Priority:
   • custom_extract_relations_prompt > custom_prompt > default prompt
   • Top-level config will be merged into sub-configs

Troubleshooting

Issue: Custom prompts not taking effect

• Check if configuration is correctly passed
• Verify field names are spelled correctly
• Check logs to confirm prompts are being read

Issue: Prompt format errors

• Ensure necessary placeholders are included
• Check if JSON format is correct
• Verify string escaping

Issue: LLM responses don't match expectations

• Adjust prompt clarity
• Add more examples
• Check LLM response format settings

Summary

Through custom prompts, you can:

• Customize extraction rules for specific domains
• Control memory update strategies
• Adjust importance evaluation criteria
• Customize graph memory operation methods

This makes powermem adaptable to various use cases and requirements.