Temporal Knowledge-Graph Memory Systems

Dated snapshot — verified 2026-08-27

Representative systems: Graphiti (open source), and Zep, which wraps it.

The architecture

Facts become edges in a graph, and every edge carries validity: when the fact became true, and when it stopped. A new fact that contradicts an old one does not overwrite it — it closes the old edge's validity interval. Queries can be asked "as of" a point in time.

What it maps to in this course

Directly onto bi-temporal modeling and supersession. If you build the Level 2 capstone as taught, you will have built a simplified version of this shape.

What to look at critically

This architecture's advantage is real and specific: it is the only common shape that answers "what did I believe then". The costs are also real — graph construction is expensive on the write path, entity resolution becomes load-bearing rather than optional, and a wrong merge corrupts a whole neighbourhood rather than one record.

Published comparisons show a substantial gap in favour of this architecture on temporal-reasoning benchmark categories specifically. That is the category where the design should win, so it is weak evidence of general superiority.