memory-systems

Use a layered architecture combining short-term session storage with long-term memory such as a vector store or knowledge graph. Store facts with entity tags and timestamps, then retrieve relevant memories just-in-time and inject them into the context window.

Vector stores provide semantic similarity search but lose relationship structure between entities. Knowledge graphs explicitly model entities and relationships, enabling traversal queries like finding products purchased by customers who bought a specific item.

Temporal knowledge graphs attach valid_from and valid_until timestamps to each fact. Queries filter facts by validity period, enabling time-travel queries and preventing outdated information from conflicting with current data.

Use a vector store when you only need semantic document retrieval without relationship reasoning. Choose a knowledge graph when queries must traverse entity relationships, and a temporal graph when facts change over time.

Memories accumulate unboundedly over time, causing retrieval to return outdated or duplicate facts. Consolidation merges related facts, updates validity periods, archives obsolete entries, and rebuilds indexes to maintain retrieval quality.