DEV Community

Ruly Altamirano
Ruly Altamirano

Posted on

LuminoraCore v1.1: Your AI Memory That Actually Travels 🧠✨

#opensource #ai #tooling #showdev

TL;DR: I built an open-source framework that gives AI persistent memory across platforms. Your ChatGPT conversations can inform Claude, your Claude chats can continue in Gemini. True data portability, privacy-first.

🎬 See It In Action (2 minutes)

Watch how memory persists when switching between ChatGPT, Claude, and other LLMs

🤔 The Problem I Was Trying to Solve

We've all been there:

  1. ChatGPT forgets your conversation from yesterday
  2. Claude doesn't know what you discussed with ChatGPT
  3. Gemini starts fresh every time
  4. Your data is locked to each platform
  5. Export options? LOL, good luck

I was frustrated. I wanted my AI to remember me, know my preferences, and travel with me across platforms.

So I built LuminoraCore.

✨ What Makes It Different

Not Just Another Chatbot Framework

LuminoraCore isn't about building yet another chatbot. It's about creating portable AI identities that persist across:

  • 🔄 Any LLM provider (OpenAI, Anthropic, DeepSeek, Groq, Mistral...)
  • 💾 Any storage backend (SQLite, PostgreSQL, MongoDB, DynamoDB...)
  • 🚀 Any application (your own apps, integrations, plugins...)

The Three Pillars

1. Persistent Memory 🧠 

from luminoracore import LuminoraCoreClient client = LuminoraCoreClient() # First conversation response = await client.send_message_with_memory( session_id="user_123", message="I'm Carlos, I love Python and basketball" ) # AI: "Nice to meet you, Carlos! Python and basketball, great combo!"  # Next day, different conversation response = await client.send_message_with_memory( session_id="user_123", message="What do you know about me?" ) # AI: "You're Carlos, you love Python and basketball!"
Enter fullscreen mode Exit fullscreen mode

How it works:

  • Automatically extracts facts from conversations (9 categories)
  • Classifies importance (0-10 scale)
  • Creates episodic memories for significant events
  • Compiles relevant context for each response

2. Affinity System 💖

The AI's relationship with you evolves naturally:

Stranger (0-10 points) → Formal, polite ↓ Acquaintance (11-30) → Friendly, helpful ↓ Friend (31-60) → Warm, personal ↓ Confidant (61+) → Very personal, empathetic
Enter fullscreen mode Exit fullscreen mode

Real example from my own usage:

  • Day 1 (Stranger): "Hello. How may I assist you today?"
  • Week 2 (Acquaintance): "Hey! How's your Python project going?"
  • Month 2 (Friend): "Carlos! Ready to talk about that basketball game?"
  • Month 6 (Confidant): "I remember you mentioned family stress last week. How are you holding up?"

The tone, depth, and personalization adapt automatically based on your relationship.

3. Provider Agnostic 🔄

Switch LLM providers mid-conversation without losing context:

# Start with ChatGPT client = LuminoraCoreClient(provider="openai") response = await client.send_message_with_memory(...) # Switch to Claude (memory persists!) client = LuminoraCoreClient(provider="anthropic") response = await client.send_message_with_memory(...) # Claude knows everything ChatGPT learned  # Try DeepSeek (20x cheaper!) client = LuminoraCoreClient(provider="deepseek") response = await client.send_message_with_memory(...) # Still remembers everything
Enter fullscreen mode Exit fullscreen mode

Supported providers:

  • OpenAI (GPT-4, GPT-3.5)
  • Anthropic (Claude 3.5, Claude 3)
  • DeepSeek
  • Groq
  • Mistral
  • Google (Gemini)
  • Local models via Ollama

🏗️ Architecture Deep Dive

How Memory Actually Works 

User: "I'm learning React" ↓ ┌─────────────────────┐ │ Fact Extraction │ ← LLM analyzes message │ (9 categories) │ └──────────┬──────────┘ │ ├─> Category: "interests" ├─> Key: "programming_framework" ├─> Value: "React" ├─> Importance: 0.7 ↓ ┌─────────────────────┐ │ Update Affinity │ ← Points +3 (shared interest) │ (0-100 scale) │ └──────────┬──────────┘ │ ├─> Total: 15 points ├─> Level: Acquaintance ↓ ┌─────────────────────┐ │ Select Personality │ ← Choose based on affinity │ (4 hierarchical) │ └──────────┬──────────┘ │ ├─> Use: "friendly_assistant" ↓ ┌─────────────────────┐ │ Compile Context │ ← Gather relevant facts │ (Dynamic selection) │ └──────────┬──────────┘ │ ├─> Include: React interest, other programming facts ↓ ┌─────────────────────┐ │ Generate Response │ ← LLM with full context │ (Provider-agnostic) │ └──────────┬──────────┘ │ ↓ Response: "That's awesome! React is great. Since you're into Python too, have you tried Django + React?"
Enter fullscreen mode Exit fullscreen mode

Fact Categories (9 Types)

  1. basics: Name, age, location, occupation
  2. preferences: Likes, dislikes, favorites
  3. interests: Hobbies, topics of interest
  4. family_friends: Relationships, social connections
  5. work: Career, projects, professional info
  6. health: Wellness, fitness, medical (sensitive)
  7. experiences: Past events, travel, life moments
  8. beliefs: Values, opinions, worldview
  9. goals: Aspirations, plans, ambitions

Episodic Memory (7 Types)

  • conversation: Significant discussions
  • achievement: Accomplishments, milestones
  • challenge: Obstacles, difficulties
  • decision: Important choices made
  • insight: Realizations, learnings
  • social: Interactions with others
  • milestone: Life events

🚀 Getting Started (5 Minutes)

Installation 

pip install luminoracore
Enter fullscreen mode Exit fullscreen mode

Minimal Example 

import asyncio from luminoracore import LuminoraCoreClient async def main(): # Initialize client  client = LuminoraCoreClient( provider="openai", # or anthropic, deepseek, etc.  api_key="your-api-key" ) # Create session  session_id = "user_carlos_session1" # First message  response = await client.send_message_with_memory( session_id=session_id, message="Hi! I'm Carlos, I'm a software engineer from Madrid" ) print(response['response']) print(f"Facts learned: {response['new_facts_count']}") # Second message (will remember Carlos)  response = await client.send_message_with_memory( session_id=session_id, message="What do you know about me?" ) print(response['response']) print(f"Total facts: {response['memory_facts_count']}") # Export all memory  export = await client.export_session(session_id) print(export) # Full JSON with all memories  asyncio.run(main())
Enter fullscreen mode Exit fullscreen mode

Output:

Response: "Hello Carlos! Nice to meet you..." Facts learned: 3 Response: "You're Carlos, a software engineer from Madrid!" Total facts: 3 { "learned_facts": [ {"category": "basics", "key": "name", "value": "Carlos"}, {"category": "basics", "key": "occupation", "value": "software engineer"}, {"category": "basics", "key": "location", "value": "Madrid"} ], "affinity": {"points": 5, "level": "stranger"} }
Enter fullscreen mode Exit fullscreen mode

💡 Real-World Use Cases

1. Personal AI Assistant 

# Remembers your preferences, schedule, habits assistant = LuminoraCoreClient(provider="anthropic") await assistant.send_message_with_memory( session_id="my_assistant", message="Schedule my morning workout" ) # AI knows: You prefer 7am, you do HIIT, 30 minutes
Enter fullscreen mode Exit fullscreen mode

2. Customer Support Bot 

# Remembers customer history, past issues, preferences support = LuminoraCoreClient(provider="deepseek") # Cheaper!  await support.send_message_with_memory( session_id=f"customer_{customer_id}", message="My order hasn't arrived" ) # AI knows: Order #12345, placed 3 days ago, shipped to Madrid
Enter fullscreen mode Exit fullscreen mode

3. Learning Tutor 

# Adapts to student's level, tracks progress tutor = LuminoraCoreClient(provider="openai") await tutor.send_message_with_memory( session_id=f"student_{student_id}", message="I don't understand recursion" ) # AI knows: Student struggles with CS concepts, prefers visual examples
Enter fullscreen mode Exit fullscreen mode

4. Mental Health Companion 

# Builds trust over time, remembers emotional patterns companion = LuminoraCoreClient(provider="anthropic") await companion.send_message_with_memory( session_id="user_therapy", message="I'm feeling anxious today" ) # AI knows: User has anxiety triggers, past coping strategies # Relationship: Confidant → very empathetic, personal response
Enter fullscreen mode Exit fullscreen mode

🎨 Advanced Features

PersonaBlend™: Mix Personalities 

# 70% friendly + 30% professional blended = client.blend_personalities( base_personality="friendly_assistant", blend_with="professional_consultant", ratio=0.7 ) response = await client.send_message_with_memory( session_id=session_id, message="Help me write a proposal", personality=blended ) # Gets: Warm tone + professional structure
Enter fullscreen mode Exit fullscreen mode

Multi-Storage Support 

# Development: SQLite (local, fast) client = LuminoraCoreClient( provider="openai", storage_backend="sqlite", db_path="./memory.db" ) # Production: PostgreSQL (robust, scalable) client = LuminoraCoreClient( provider="anthropic", storage_backend="postgresql", connection_string="postgresql://..." ) # Cloud: DynamoDB (managed, serverless) client = LuminoraCoreClient( provider="deepseek", storage_backend="dynamodb", table_name="luminora_memories" )
Enter fullscreen mode Exit fullscreen mode

Export & Migrate 

# Export from ChatGPT setup export = await client_gpt.export_session("user_123") # Import to Claude setup client_claude = LuminoraCoreClient(provider="anthropic") await client_claude.import_session("user_123", export) # All memories preserved!
Enter fullscreen mode Exit fullscreen mode

🎯 Roadmap

v1.2 (Q1 2025): Semantic Search 

# Instead of exact key matching results = await client.semantic_search( user_id="carlos", query="what sports does carlos like", # Natural language!  top_k=5 ) # Returns: basketball, running, cycling...
Enter fullscreen mode Exit fullscreen mode

v1.3 (Q2 2025): Knowledge Graphs 

# Connect related facts graph = await client.get_knowledge_graph(user_id="carlos") # Carlos → knows → John → plays → basketball # ↓ # likes → basketball
Enter fullscreen mode Exit fullscreen mode

v1.4 (Q3 2025): Memory Compression 

# Problem: 6 months = 92K tokens = $2,775 per request! # Solution: Tiered compression (recent=full, old=compressed) # Result: 20K tokens = $600 per request (78% savings)
Enter fullscreen mode Exit fullscreen mode

v2.0 (2026): API SaaS

  • Managed API service
  • Multi-tenant
  • Integrations (LangChain, N8N, etc.)
  • Browser extension marketplace

🤝 Contributing

LuminoraCore is MIT licensed and community-driven.

Ways to contribute:

  1. 🐛 Report bugs
  2. 💡 Suggest features
  3. 📝 Improve documentation
  4. 🧪 Add test cases
  5. 🔧 Submit PRs

Priority areas:

  • Semantic search implementation
  • Additional LLM providers
  • Storage backend optimizations
  • Real-world use case examples

📚 Resources

Code & Docs:

🎬 Try It Now 

# Install pip install luminoracore # Clone examples git clone https://github.com/luminoracore/luminoracore cd luminoracore/examples # Run simple chat python simple_chat.py
Enter fullscreen mode Exit fullscreen mode

🙏 Why I Built This

I'm tired of:

  • 😤 AI platforms that forget me
  • 😤 Vendor lock-in
  • 😤 Opaque costs
  • 😤 Lost conversation history

I wanted:

  • ✅ My data, my control
  • ✅ Privacy-first (self-hosted option)
  • ✅ Portable (works anywhere)
  • ✅ Transparent (open source)

If you feel the same, give it a try ⭐

💬 Let's Talk

What would you build with persistent AI memory?

  • Personal assistant?
  • Customer service bot?
  • Learning tutor?
  • Something else?

Drop a comment below! I'd love to hear your ideas and help you build it.

P.S. If you found this useful, consider:

Building in public. Let's make AI memory better together. 🚀

Updated: November 2025 | Version: 1.1 | Author: @rulyaltamirano

Top comments (1)

Collapse
 
avanichols_dev profile image
Ava Nichols •

Love this idea of portable AI identity and the affinity system! How are you thinking about encrypting/sandboxing sensitive categories like health or beliefs across different storage backends?