memory-systems

Implement layered memory: working memory in the context window, short-term session storage, and long-term persistence via vector stores or knowledge graphs. Store facts with entity tags and timestamps, then retrieve relevant memories just-in-time for each task.

Vector RAG retrieves semantically similar documents but loses relationship structure between entities. Knowledge graphs explicitly model entities and relationships, enabling traversal queries like "what products did customers who bought X also buy" that vector stores cannot answer.

Use temporal knowledge graphs when facts change over time and you need time-travel queries, such as reconstructing a user's address on a specific date. They add valid_from and valid_until timestamps to facts, preventing outdated information from conflicting with current data.

Vector stores embed facts as isolated chunks without preserving connections between entities. Similarity search can retrieve a directly stated fact but cannot traverse relationships, so multi-hop questions across entities return incomplete or wrong results.

Run periodic memory consolidation: identify duplicate facts grouped by subject and predicate, merge related entries keeping the highest-confidence version, update validity periods, and archive obsolete facts. Trigger consolidation after significant accumulation or on a schedule.