🚀 The feature

TLDR: Add a get_profile() method that generates an evolving, compact user profile (200–400 tokens) for use as baseline session context. Profiles update automatically from memories and provide a semantic “who the user is” layer alongside episodic RAG retrieval.

I want to add a get_profile() method that generates a continuously-updated user profile for every AI interaction. This could piggy-back off of custom fact extraction methods.

• Compact (200–400 tokens)
• Configurable (prompt, size, update rules)
• Cached + auto-updated
• Complements RAG

Example Config (mirrors custom_fact_extraction_prompt)

from mem0 import Memory

config = {
    "llm": {"provider": "openai", "config": {"model": "gpt-4o-mini"}},

    # NEW: Custom profile prompt
    "custom_profile_prompt": """
        Generate a concise user profile (200-400 tokens) covering:
        - Demographics and role
        - Communication preferences
        - Key expertise/skills
        - Major interests and preferences
        - Current goals/context
        Focus on stable, high-level traits, not specific events.
    """,

    # NEW: Profile auto-update config
    "profile_config": {
        "max_tokens": 400,
        "auto_update": True,
        "update_triggers": {
            "memory_count": 10,     # Update after N new memories
            "time_elapsed": 86400   # Update daily
        }
    }
}

m = Memory.from_config(config_dict=config)

# Profile auto-updates in background (like custom facts)
m.add(messages, user_id="alice")

# Fast retrieval - always returns cached version
profile = m.get_profile(user_id="alice")

Usage in session context

system_prompt = f"""You are an AI assistant.

User Profile:
{profile}

Relevant Memories:
{m.search(query=user_message, user_id="alice", limit=3)}
"""

Motivation, pitch

Mem0 currently excels at extracting episodic memories (facts/events) via RAG. But there’s no way to generate a holistic, semantic profile that answers “who is this person?” Humans maintain both episodic (specific) and semantic (general) memory; mem0 only covers episodic well today.

My Use Case (I'm building MindMe, an AI mental health companion)

In MindMe (AI mental health assistant), each session would includes:

• User Profile (300 tokens, every session)
• Retrieved Memories (200 tokens, query-specific)
• Conversation History Window(last 5 turns)

Example profiles:
Customer Support: "Alice is a premium customer, prefers technical details, timezone PST"
Healthcare: "Patient has diabetes, prefers morning appointments, anxious about needles"
Education: "Student struggles with calculus, visual learner, preparing for AP exam"