Why temporal knowledge graphs need two clocks
A temporal knowledge graph must distinguish database history from world history. Transaction or recorded time establishes when an assertion entered the system. Valid time describes the interval during which the assertion applies to the modeled domain.
Without both, late-arriving corrections and retrospective facts become ambiguous. An append-only operation history can preserve recorded time, while fact nodes and relationships carry valid intervals and explicit supersession.
Represent change without destructive updates
Keep source episodes immutable. Model new interpretations as new facts and connect them using typed relationships such as Supersedes, Retracts, Contradicts, and SupportedBy. A current-truth projection can select the active interpretation while historical queries retain the complete path.
This pattern supports audit and reproducibility, but it needs compaction and erasure workflows. Append-only should describe the logical history, not imply that every physical object remains forever.
Define current, valid-at, and recorded-as-of queries
A temporal query contract should state which clock it uses. Current selects the latest admitted interpretation. Valid-at selects claims whose valid interval contains a time. Recorded-as-of reconstructs what the system knew at a past point.
For ambiguous or conflicting claims, return multiple candidates with status and evidence rather than silently collapsing them.
Keep temporal indexes derived and repairable
Shard-local SQL indexes can efficiently select valid intervals and reverse incidence. Cross-shard and semantic indexes are derived projections with explicit watermarks. The authoritative operation log remains the recovery source.
A semantic digest must cover schema, merge policies, expiry rules, unresolved dependencies, and all query-visible state—not just atom identifiers—otherwise equal digests can mask different behavior.
Use temporal context where change is a feature
Useful domains include agent memory, investigations, customer state, evolving organizational knowledge, configuration history, identity and entitlement lineage, scientific claims, and incident reconstruction.
A temporal graph is unnecessary when all data is static, relationships are shallow, and ordinary transactional records answer the questions. Start with the simplest model that preserves the required truth semantics.