Memory Blossom

An advanced cognitive architecture for AI assistants that combines multi-layered memory systems with adaptive criticality assessment.

Overview

Memory Blossom is a sophisticated system designed to enhance AI assistants with human-like memory organization and optimal response generation. It works by:

1. Storing memories in specialized categories using neurally-inspired embedding models
2. Dynamically retrieving contextually relevant memories during conversations
3. Assessing response quality across order-chaos dimensions
4. Adaptively tuning responses to find an optimal "sweet spot" between predictability and novelty

Theory: Why Memory Blossom Matters

This architecture addresses two key limitations in current AI systems:

1. The Memory Problem

Standard large language models lack true episodic memory. They operate primarily on their training data and immediate conversation context, without building persistent, contextually-aware memories over time.

Memory Blossom implements a multi-layered memory system inspired by human cognitive psychology:

• Explicit Memory: Factual knowledge and information
• Emotional Memory: Affective experiences and feelings
• Procedural Memory: Skills and know-how
• Flashbulb Memory: Vivid, identity-defining experiences
• Somatic Memory: Sensory and physical experiences
• Liminal Memory: Emerging ideas and possibilities
• Generative Memory: Creative and imaginative concepts

Each memory type uses a specialized neural embedding model to optimize semantic representation specific to that category of information.

2. The Criticality Problem

AI responses often fall into one of two suboptimal zones:

• The Ordered Zone: Responses that are coherent but predictable/generic
• The Chaotic Zone: Responses that are novel but potentially incoherent

Memory Blossom implements a "criticality assessment" system that evaluates responses along three dimensions:

1. Statistical Likelihood: How predictable is the response based on language model probabilities?
2. Semantic Novelty: How different is the response from the input query?
3. Internal Consistency: How coherent is the response's internal structure?

This assessment helps identify the "Critical Zone" - where responses are both coherent and novel - and adjusts generation parameters to reach this optimal state.

Components

memory_blossom.py

This file implements the core memory system:

• Memory class: Stores individual memories with rich metadata
• MemoryBlossom class: Manages memory categorization and retrieval
• Specialized embedding models: Seven different neural embedding models for specific memory types
• Dynamic classification: Uses GPT-3.5-Turbo to classify memories (with fallback heuristics)
• Context-aware retrieval: Analyzes conversation flow to prioritize relevant memory types
• Persistence: Save/load functionality for long-term memory retention

chat.py

This file implements the conversational agent:

• MemoryBlossomChatbot class: Integrates Memory Blossom with OpenAI's API
• Criticality assessment: Evaluates response quality along order-chaos spectrum
• Adaptive regeneration: Adjusts temperature parameter to reach optimal response zone
• Command-line interface: Simple REPL for interactive testing
• System prompt injection: Incorporates relevant memories into prompts

Usage

1. Install dependencies:

pip install sentence-transformers sklearn numpy openai python-dotenv

2. Set up your OpenAI API key:

export OPENAI_API_KEY="your-api-key"

3. Run the interactive chat:

python chat.py

Command Line Interface

The simple CLI supports:

• Normal chat interaction
• memories command to view recent memories
• clear command to reset conversation context
• exit command to quit

Extensions and Customization

• Add custom memory types by extending the memory_stores dictionary
• Implement more sophisticated emotion scoring by replacing the keyword-based approach
• Use different LLMs by changing the OpenAI model parameter
• Adjust criticality thresholds to tune the sweet spot between novelty and coherence

References & Inspiration

This system draws inspiration from:

• Cognitive psychology models of human memory
• Complex systems theory, particularly self-organized criticality
• Neural embedding techniques for semantic representation
• LLM prompt engineering best practices

Requirements

• Python 3.8+
• Recommended: 32GB+ RAM (for multiple embedding models)
• Optional: NVIDIA GPU with 4GB+ VRAM
• Internet connection for OpenAI API calls