Bitemporal Graph Memory Store for Conversational AI
A recently developed memory system for conversational AI agents employs a Neo4j property graph featuring HNSW vector indexes alongside a comprehensive bitemporal data model. Each memory is represented as an immutable identity node connected to versioned content nodes that utilize two closed-open time intervals: valid time (the period during which the fact was accurate) and transaction time (the moment it was documented). This architecture facilitates point-in-time semantic retrieval while preserving historical data, thus preventing context window fatigue or exposure to external data. Semantic connections between memories are upheld through cosine similarity applied to 1024-dimensional embeddings.
Key facts
- Memory store uses Neo4j property graph
- HNSW vector indexes are used
- Bitemporal data model with valid time and transaction time
- Memories are immutable identity nodes with versioned content
- Supports point-in-time semantic retrieval
- Avoids context window exhaustion
- Avoids sending personal data to third parties
- Semantic edges maintained via cosine similarity on 1024-dimensional embeddings
Entities
—