Knowledge graph memory stores facts as linked entities and relationships; vector memory stores text chunks as embeddings and finds them by similarity.
A knowledge graph represents 'Alice manages the billing service' as two nodes and an edge, often with timestamps. Queries can walk the graph: who else does Alice manage, what changed last month. Zep's temporal graph and Mem0's Pro-tier graph memory take this approach, and Basic Memory builds a graph over Markdown links. Vector memory instead keeps the sentence as text, embeds it and returns it when a similar question arrives. It is simpler and needs no schema.
Each has a natural fit. Graphs are strong when relationships and time matter and when you need to answer 'what is true now' precisely. Vectors are strong for unstructured documents, fuzzy questions and getting started quickly. Graphs require extraction (a model deciding what the entities are), which introduces its own errors. Many products combine both: a vector index over source text, plus structured notes or a graph over the facts extracted from it.