Temporal degradation
Every triple you extracted last year sits in your graph with the same epistemic weight as the triple you extracted this morning. The graph has no concept of belief decay. It cannot distinguish between a fact and a fossil.
Temporality in knowledge graphs is almost always conflated with timestamps. A created field on a node is however not temporal modeling. It records when the assertion entered the system. It says nothing about the interval during which the assertion is valid, nor about the rate at which your confidence in it should erode.
Bitemporal modeling has existed for a long time, yet almost universally absent from production. The deeper problem is epistemic decay. Expiration assumes you know when a fact stops being true. You rarely do. What you actually have is a credence (a degree of belief) that decreases as a function of time, domain volatility, and source reliability. I’ll refrain talking about Bayesian networks but do look it up.
All of this is really a prior that should propagate through your graph via confidence-weighted traversals. The consequence for GraphRAG is a retrieval safety problem. Stale nodes with high historical centrality accumulate epistemic authority they no longer deserve. When your retrieval layer surfaces context for an LLM, it preferentially returns high-centrality nodes. If those nodes encode the organizational reality of three years ago, your agent reasons with great confidence from a false prior.
This is historical hallucination in the graph in a way, laundered through retrieval, and then likely blamed on the LLM.
What temporal degradation requires in practice is an edge-level decay functions, parameterized per ontology class, not global, not uniform. Confidence propagation on traversal, not just on storage. Re-extraction triggers when node confidence drops below a domain-specific threshold. And more importantaly a query API that exposes temporal confidence as a first-class return value, not a hidden metadata field that analysts never see. The graph should be able to tell you: I believe this, but I believed it more strongly two years ago.
Until it can, you don’t have a knowledge graph (or agent memory or context graph). You have an append-only log with a SPARQL/GQL endpoint.