memory-systems

To persist agent memory across sessions, you need a memory architecture that supports working, short-term, long-term, and knowledge-graph layers. This enables agents to maintain continuity and reason over past cross-session interactions.

A temporal knowledge graph for agent entity tracking is a memory layer that records how facts and relationships evolve over time. It enables temporal reasoning by applying validity timestamps to stored entity memories.

To retrieve and consolidate memories by entity, you query the memory architecture using semantic search and entity-based filters. Consolidation patterns then merge relevant working, short-term, and long-term memories into the active context.

Yes, numpy is required as a dependency to implement this cross-session memory architecture. It provides the numerical processing foundation needed for semantic search retrieval and memory consolidation operations.

The best way to structure agent memory for customer support is a layered architecture combining working, short-term, long-term, and entity memory. This allows the agent to remember user preferences across chats and tailor responses effectively.

You should use a knowledge graph layer instead of simple long-term memory when your agent requires temporal reasoning about how facts and relationships evolve over time, rather than just retrieving static historical context.