Reference

Bi-temporal knowledge graph

Behavior

LongMemEval temporal check	Result	Meaning
Gold-session provenance visible at question time	0.936	Some gold session IDs are absent or timestamp-incompatible in the cleaned records
Future-session exclusion	1.000	The bitemporal filter excluded every session dated after the question

The reported value measures visibility semantics. Model answer quality has its own measure.

Example bitemporal difference

Query	Valid at requested time	Known at requested time	Returned
Historical state before correction arrived	Yes	Yes	Original fact
Same valid date, knowledge after correction	Yes	No for original revision	Corrected fact
Timestamp equals half-open interval end	No	n/a	Excluded

TemporalFact tracks valid time (when a claim applies) and transaction time (when the system knows it), supporting historical queries and late corrections.

How it works

An assertion is append-only and carries two intervals. A correction learned today requires the host to close an older assertion’s transaction interval. close_transaction returns a replacement value for that operation. Its historical valid interval remains in the record. Query at filters valid time. Query known_at filters transaction time. Each interval must contain its requested timestamp. Contradictions create explicit edges or superseding revisions. Prior revisions stay available.

The temporal functions operate on supplied records. The host persists history and performs correction writes atomically. Use timezone-aware timestamps and half-open intervals [start, end). An open end means the interval continues indefinitely. See temporal and provenance utilities for joins and shared revision lineage.

Research basisZep uses a temporally aware graph to maintain historical relationships for agent memory. The temporal knowledge-graph survey catalogs representations and inference tasks for facts that change over time. Mari adds explicit valid-time and transaction-time query semantics. Applications choose interval boundaries and contradiction policies for their domain.

valid timeJanAug

transaction timelearned Sep 01onward

Half-open valid-time and transaction-time queries

from datetime import datetime, timezone

from mari_kit.graph import TemporalFact, query_temporal_facts

utc = timezone.utc
facts = [
    TemporalFact(
        fact_id="refund-window@1",
        subject="plan:enterprise",
        predicate="refund_window_days",
        object=30,
        valid_from=datetime(2026, 1, 1, tzinfo=utc),
        valid_to=datetime(2026, 9, 1, tzinfo=utc),
        recorded_at=datetime(2026, 1, 3, tzinfo=utc),
    )
]

visible = query_temporal_facts(
    facts,
    at=datetime(2026, 6, 1, tzinfo=utc),
    known_at=datetime(2026, 8, 1, tzinfo=utc),
)