Graph Database Entities Not Utilized in Recall within

[pihang]

🐛 Describe the bug

memo/proxy/main.py
It seems that when answering questions in memo/proxy/main.py, the recall mechanism doesn't utilize the dictionary information from the graph database.

Current code:
memories_text = "\n".join(memory["memory"] for memory in relevant_memories["results"])

In memo/memory/main.py, the memory search actually includes parallel searches in both vector storage and graph storage. The code returns:
return {"results": original_memories, "relations": graph_entities}

However, here only original_memories from vector storage recall is handled, and the graph database entities (graph_entities) are not used in constructing the memory recall.

How should we handle the recall of entities from the graph database?

For instance, data structured as follows is not being utilized:
search_results.append({"source": item[0], "relationship": item[1], "target": item[2]})
Example:

{
  "memories": [...],
  "entities": [
    {"source": "peter", "relation": "identity", "destination": "spiderman"}
  ]
}

[Cirr0e]

Here's a proposed solution to integrate graph database entities into the memory recall process:

def enhanced_memory_recall(relevant_memories):
    # Combine vector storage memories
    memories_text = "\n".join(memory["memory"] for memory in relevant_memories["results"])
    
    # Add graph entity information to provide additional context
    if "relations" in relevant_memories and relevant_memories["relations"]:
        # Format graph entities into a readable string
        graph_context = "\n".join([
            f"Relation: {entity.get('relationship', 'Unknown')} "
            f"From: {entity.get('source', 'Unknown')} "
            f"To: {entity.get('target', 'Unknown')}"
            for entity in relevant_memories.get("relations", [])
        ])
        
        # Combine memories with graph entity context
        full_context = f"{memories_text}\n\n--- Related Entities ---\n{graph_context}"
    else:
        full_context = memories_text
    
    return full_context

Key improvements:

1. Preserves existing vector memory retrieval
2. Adds graph entity information if available
3. Creates a comprehensive context string
4. Provides optional fallback if no graph entities exist

Recommended modifications in memo/proxy/main.py:

# Replace existing memory text generation
memories_text = enhanced_memory_recall(relevant_memories)

Rationale and References:

• Based on the Mem0.ai documentation on Graph Memory integration [Citation: Mem0.ai Documentation]
• Solves the issue of unused graph database entities
• Provides a flexible approach to incorporating relational context

Additional Recommendations:

1. Ensure your Neo4j configuration is up to date (version v1.1+)
2. Validate graph entity extraction in your memory creation process
3. Consider adding logging to track graph entity utilization

Potential Enhancement:
For more advanced use cases, you might want to develop a more sophisticated entity integration strategy that can:

• Weight graph entity importance
• Handle more complex relationship types
• Provide more nuanced context generation