Semantic Memory

SEMANTIC MEMORY TYPES: 1. FACTUAL KNOWLEDGE: - "Python was created by Guido van Rossum" - "The Earth orbits the Sun" - "HTTP status 404 means 'Not Found'" 2. CONCEPTUAL KNOWLEDGE: - "Inheritance is a type of code reuse" - "APIs enable system communication" - "Recursion involves self-reference" 3. RELATIONAL KNOWLEDGE: - "Python IS-A programming language" - "FastAPI USES Pydantic" - "Transformers IMPROVED-UPON RNNs" --- KNOWLEDGE REPRESENTATION: FLAT (Simple key-value): { "python_creator": "Guido van Rossum", "python_year": 1991, "python_type": "programming language" } HIERARCHICAL (Taxonomy): Programming Languages ├── Compiled │ ├── C │ ├── Go │ └── Rust └── Interpreted ├── Python ├── JavaScript └── Ruby GRAPH (Knowledge graph): ┌─────────┐ created_by ┌───────────────┐ │ Python │─────────────→│ Guido van │ └────┬────┘ │ Rossum │ │ └───────────────┘ │ is_a ▼ ┌─────────────────────┐ │ Programming Language │ └──────────┬──────────┘ │ has_feature ▼ ┌─────────────────────┐ │ Dynamic Typing │ └─────────────────────┘ VECTOR (Embeddings): "Python" → [0.23, -0.45, 0.12, ..., 0.89] "JavaScript" → [0.21, -0.42, 0.15, ..., 0.85] (Similar concepts = similar vectors)

SEMANTIC MEMORY SOURCES IN LLMs: 1. PARAMETRIC MEMORY (Built-in): - Knowledge encoded in model weights - Learned during pre-training - Fixed after training (can't update) LLM: "Python was created in 1991" (Knows this from training data) 2. NON-PARAMETRIC MEMORY (External): - Knowledge bases, databases - Retrieved at inference time - Can be updated independently RAG: [Query] → [Retrieve docs] → [Augment prompt] (Gets current info from external source) --- COMBINING MEMORY TYPES: User: "What's the latest Python version?" Parametric (may be outdated): "Python 3.11 is the latest" (training cutoff) + Non-parametric (current): [Retrieved: "Python 3.12 released Oct 2023"] Combined answer: "Python 3.12 is the current stable release." --- SEMANTIC MEMORY SYSTEM ARCHITECTURE: ┌─────────────────────────────────────────────┐ │ USER QUERY │ └──────────────────┬──────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────┐ │ QUERY UNDERSTANDING │ │ Extract: entities, intent, concepts │ └──────────────────┬──────────────────────────┘ │ ┌──────────┴──────────┐ ▼ ▼ ┌───────────────┐ ┌───────────────────────┐ │ KNOWLEDGE │ │ VECTOR │ │ GRAPH │ │ DATABASE │ │ │ │ │ │ Structured │ │ Semantic similarity │ │ relationships │ │ search │ └───────┬───────┘ └───────────┬───────────┘ │ │ └──────────┬─────────────┘ ▼ ┌─────────────────────────────────────────────┐ │ KNOWLEDGE FUSION │ │ Combine retrieved facts + relationships │ └──────────────────┬──────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────┐ │ LLM GENERATION │ │ Generate response using fused knowledge │ └─────────────────────────────────────────────┘ KNOWLEDGE UPDATE STRATEGIES: - Vector DB: Add new embeddings - Knowledge graph: Add new triples - Parametric: Requires fine-tuning (expensive)