Agent Memory Is Only as Good as Its Schema

When using knowledge graphs for AI agent memory, letting the LLM freely decide entity types and relationship labels produces generic, unqueryable graphs where everything becomes a 'Topic' with 'RELATES_TO' edges. Zep's Graphiti library solves this by letting developers define a typed ontology using Pydantic models — custom entity types (e.g., Project, Technology) and edge types (e.g., WORKS_ON, USES_TECHNOLOGY) with structured attributes and source/target constraints. The extraction pipeline then classifies against these definitions rather than guessing. This enables multi-hop reasoning that flat vector retrieval can't support, typed filtering by attributes, and bi-temporal tracking of facts. Zep enforces a 10/10/10 limit (10 entity types, 10 edge types, 10 fields each) to keep schemas focused. Context templates then assemble typed, temporally annotated facts into prompt-ready blocks for agents.