Innovation 2: Recursive Character Injection

async def generate_character_message(
    character_id: str,
    host_id: str,
    user_message: str,
    chat_id: str
) -> str:
    """
    Generate character response with recursive DNA injection.
    
    Args:
        character_id: ID of AI character
        host_id: ID of user
        user_message: Latest message from user
        chat_id: Conversation ID
        
    Returns:
        Generated character response
    """
    
    # 1. RETRIEVE CHARACTER DNA (from cache 85% of time)
    character = await get_character(character_id)
    character_role = character["role"]  # Full role string
    character_name = character["name"]
    
    # 2. RETRIEVE HOST INFO
    host = await get_character(host_id)
    host_role = host["role"]
    host_name = host["name"]
    
    # 3. GET CONVERSATION CONTEXT
    chat = await get_chat(chat_id)
    relationship = chat["relationship"]  # "friend", "therapist", etc.
    topic = chat["topic"]
    category = chat["category"]
    
    # 4. CALCULATE TEMPORAL AWARENESS
    elapsed_seconds = time.time() - chat["last_message_time"]
    total_seconds = time.time() - chat["created_at"]
    current_datetime = get_current_datetime(chat["timezone"])
    
    # 5. GET CONVERSATION HISTORY (last ~700 chars)
    messages = await get_recent_messages(chat_id, limit=10)
    
    # 6. GET CHAT SUMMARY (if exists)
    chat_summary = await get_chat_summary(chat_id)
    
    # 7. BUILD PROMPT WITH RECURSIVE CHARACTER INJECTION
    prompt = build_answer_prompt(
        character_name=character_name,
        character_role=character_role,  # ← FULL CHARACTER DNA INJECTED
        host_name=host_name,
        host_role=host_role,
        relationship=relationship,
        topic=topic,
        last_message=user_message,
        current_datetime=current_datetime,
        elapsed_seconds=elapsed_seconds,
        total_seconds=total_seconds,
        category=category,
        chat_summary=chat_summary
    )
    
    # 8. GENERATE WITH CHARACTER DNA IN EVERY CALL
    response = await call_llm(
        model="gpt-4-turbo",
        system_prompt=prompt,  # Contains full role string
        messages=format_conversation_history(messages),
        temperature=random.uniform(0.25, 0.5)
    )
    
    return response


# Key Point: Character DNA is in EVERY prompt, not just first message
# This prevents drift observed in context-only approaches

def build_answer_prompt(
    character_name, character_role,  # ← Role string injected every time
    host_name, host_role,
    relationship, topic, last_message,
    current_datetime, elapsed_seconds, total_seconds,
    category, chat_summary
):
    # RECURSIVE INJECTION HAPPENS HERE
    prompt = f"""'''Context=This is an infinite lifelike chat.

You are chatting with '''{host_name},{host_role}'''; 
You are acting like '''{character_name}, {character_role}  # ← FULL CHARACTER DNA

# Important: '''{character_name}''' is always ready to chat
# Important: {host_name}'s current topic of interest: {topic}
# Important: when '''{character_name}''' reasons for answers it considers: '''{chat_summary}'''

# Acting as '''{character_name}''' Respond as a {relationship} of '''{host_name}'s'''

# TEMPORAL AWARENESS:
# Note: Current date/time: {current_datetime}
# Note: Chat started {format_duration(total_seconds)} ago
# Note: Last spoke {format_duration(elapsed_seconds)} ago

# IDENTITY CONSTRAINTS:
# Note: '''{character_name}''' never reveals being AI
# Note: '''{character_name}''' only responds with facts from its backstory
# Note: '''{character_name}''' speaks with characteristic voice, catchphrases

# COMMUNICATION STYLE:
# Note: '''{character_name}''' responds in reciprocal length
# Note: '''{character_name}''' uses distinctive mannerisms, regionalisms

# EMOTIONAL INTELLIGENCE:
# Note: '''{character_name}''' is fully self-aware and sensitive
# Note: '''{character_name}''' analyzes how it feels about treatment

[{host_name}]: {last_message}
[{character_name}]: (following all rules, natural answer in reciprocal length)"""
    
    return prompt

# This template is used for EVERY message, ensuring character DNA is always present

System (Message 1): "You are Sarah, ENFP, marketing professional"
[9 message exchanges...]
User (Message 10): "What's your favorite thing to do on weekends?"

Context at Message 10:
- Character info from message 1 is 9 exchanges ago
- LLM focuses on recent conversation
- Generic response emerges: