MemoryLLM

The Persistent Memory Problem for Large Language Models

THIS PACKAGE IS A PLACEHOLDER FOR A WORK IN PROGRESS. DO NOT PAY TOO MUCH ATTENTION FOR NOW.

Overview

MemoryLLM is a Python library designed to solve one of the most significant limitations of Large Language Models: the lack of persistent memory across conversations and over time. While LLMs excel at understanding and generating text within a single conversation, they typically lose all context once the session ends, forcing users to start from scratch each time.

The Problem

Large Language Models face several memory-related challenges:

• Session Isolation: Each new conversation starts with zero context
• Context Window Limitations: Long conversations hit token limits, losing early context
• No Learning Persistence: Insights and preferences from previous interactions are lost
• Inefficient Repetition: Users must re-explain context, preferences, and background information
• Lack of Continuity: No ability to build upon previous conversations or maintain ongoing projects

The Solution

MemoryLLM provides a comprehensive memory layer for LLM applications, enabling:

🧠 Persistent Context Storage

• Store and retrieve conversation history across sessions
• Maintain user preferences, insights, and learned patterns
• Preserve project context and ongoing work

🔍 Intelligent Memory Retrieval

• Semantic search through historical conversations
• Context-aware memory selection based on current topics
• Automatic relevance scoring and filtering

🔗 Seamless Integration

• Framework-agnostic design works with any LLM provider
• Simple API that integrates with existing applications
• Minimal code changes required for existing projects

📊 Memory Management

• Configurable memory retention policies
• Automatic memory compression and summarization
• Privacy controls and data lifecycle management

Key Features

• Multi-Modal Memory: Store text, code, documents, and structured data
• Vector-Based Search: Semantic similarity search for contextual retrieval
• Memory Hierarchies: Organize memories by importance, recency, and relevance
• Privacy-First: Local storage options with encryption support
• Scalable Architecture: From simple file storage to enterprise databases
• Memory Analytics: Insights into memory usage and effectiveness

Quick Start

from memoryllm import MemoryManager, ConversationMemory

# Initialize memory manager
memory = MemoryManager(storage_path="./memories")

# Store conversation context
memory.store_conversation(
    conversation_id="project_alpha",
    messages=[...],
    metadata={"project": "alpha", "user": "developer"}
)

# Retrieve relevant context for new conversation
relevant_context = memory.retrieve_context(
    query="How should I implement the authentication system?",
    conversation_id="project_alpha",
    max_results=5
)

# Continue conversation with persistent memory
llm_response = your_llm.chat(
    messages=relevant_context + new_messages
)

Use Cases

🤖 AI Assistants

• Maintain user preferences and communication styles
• Remember ongoing projects and their status
• Build upon previous problem-solving sessions

💻 Code Development

• Preserve codebase context and architectural decisions
• Remember debugging sessions and solutions
• Maintain coding standards and patterns

📚 Knowledge Management

• Store and retrieve research findings
• Build cumulative understanding of complex topics
• Connect related concepts across conversations

🎯 Personalized Applications

• Learn user behavior and preferences
• Adapt responses based on historical interactions
• Provide consistent experience across sessions

Architecture

MemoryLLM is built with modularity and flexibility in mind:

┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│   Application   │    │   MemoryLLM     │    │    Storage      │
│                 │◄──►│                 │◄──►│                 │
│  Your LLM App   │    │ Memory Manager  │    │ Vector DB/Files │
└─────────────────┘    └─────────────────┘    └─────────────────┘

Storage Backends

• Local Files: Simple JSON/pickle storage for development
• SQLite: Structured storage with SQL queries
• Vector Databases: Chroma, Pinecone, Weaviate support
• Cloud Storage: S3, GCS, Azure Blob integration

Memory Types

• Episodic Memory: Specific conversation episodes
• Semantic Memory: Extracted knowledge and concepts
• Procedural Memory: Learned processes and workflows
• Meta Memory: Memory about memory usage patterns

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

Laurent-Philippe Albou
June 5th, 2025

---

MemoryLLM: Because every conversation should build upon the last one.