Memory Engine

Overview

The Memory Engine is MetaMemory's central orchestration layer. It manages the full lifecycle of a memory, from creation through retrieval to eventual consolidation, coordinating across multiple subsystems.

Memory Lifecycle

When you create a memory, the engine performs the following steps automatically:

• CRUD storage: persists the memory content and metadata to PostgreSQL via Prisma
• Multi-vector encoding: generates embeddings across four specialized vector spaces (semantic, emotional, process, context)
• Emotional tagging: extracts and encodes emotional state as continuous trajectories
• Graph relationships: identifies entities and creates relationship edges in Neo4j (if configured)
• Episode tracking: associates the memory with an active episode if one exists

Five-Channel Retrieval

When you search, the engine runs five parallel retrieval channels and fuses their results via weighted Reciprocal Rank Fusion (RRF):

• Semantic: cosine similarity on 1536d OpenAI embeddings
• Temporal: recency-weighted scoring with exponential decay
• Emotional: trajectory-level similarity matching
• Keyword: BM25-style lexical matching
• Graph: entity relationship traversal via Neo4j

The fusion weights are adaptive. They start at configured defaults and shift based on retrieval feedback through the adaptive retrieval system.

Initialization

import { MemoryEngine } from 'metamemory';

const engine = new MemoryEngine({
  // All configuration options are optional
  similarityThreshold: 0.55,
  recencyDecayDays: 365,
});

await engine.initialize();

The initialize() call connects to all configured backends (PostgreSQL, Pinecone, Redis, Neo4j) and loads any existing strategy state.