Temporal Knowledge Graphs for Agent Memory: State of the Art and Production Tradeoffs (2026)

Foundational Systems:
- TLogic (AAAI 2022) pioneered explainable temporal reasoning via mined logical rules from temporal random walks. Its influence persists through GenTKG and TILP (2024).
- TANGO (2021) introduced continuous-time modeling via Neural ODEs, enabling reasoning at arbitrary time points rather than discrete snapshots.
- RE-GCN (2021) established the autoregressive paradigm (R-GCN + GRU over snapshots). It has been extended by TRCL (2025, +contrastive learning, 45.07% MRR on ICEWS14) and MTS-RE-GCN (February 2026, +14.6% MRR improvement with multi-task spatial reasoning).

Time-Decaying Relevance – the core question:
- TimeDE (IJCNN 2024) formalizes time decay via multivariate Hawkes processes with attention-based heterogeneous fact scoring.
- DynTKG (July 2025) combines time-decay Hawkes processes with dynamic subgraph pruning and causal-aware knowledge distillation, achieving 55.26% MRR on ICEWS05-15.
- Agent memory systems have largely favored discrete validity models (Zep’s bi-temporal timestamps with valid/invalid edges) over continuous exponential decay, because they support temporal queries, maintain audit trails, and are more interpretable.

Agent Episodic Memory – the most active frontier (2025-2026):
- Zep/Graphiti (January 2025): Three-tier temporal KG (episode/semantic/community subgraphs) with bi-temporal data model. Achieves 94.8% on DMR, +18.5% on LongMemEval, +38.4% on temporal reasoning questions specifically.
- MAGMA (January 2026): Four orthogonal graphs (temporal, causal, semantic, entity) with dual-process memory. Achieves 0.700 on LoCoMo, 61.2% on LongMemEval – outperforming all competitors.
- AriGraph (IJCAI 2025): Joint episodic-semantic memory with continuous outdated knowledge detection.

Critical benchmark warning: The 2026 TKG Evolution Benchmark revealed that standard ICEWS/GDELT benchmarks can be solved by trivial co-occurrence heuristics. New benchmarks with bias-correction and a novel “knowledge obsolescence prediction” task show existing methods achieving below 15% Recall@50.

The report contains 40+ citations from 2024-2026 with full source URLs.

Production Blueprint

This topic is high impact because temporal reasoning and fact lifecycle management in agent memory graphs directly determines whether an agent system remains reliable under scale, turnover, and policy change. Teams that treat this as a one-time architecture choice usually accumulate hidden risk in retrieval quality, observability, or governance controls. The safer pattern is to treat memory design as an operating discipline with explicit gates, measurable outcomes, and rollback paths.

Technical Gates Before Launch

• Choose a time model explicitly: valid-time only, transaction-time only, or full bi-temporal storage with retroactive correction support.
• Define obsolescence rules per relation type so outdated facts are invalidated deterministically instead of silently competing at retrieval time.
• Benchmark temporal queries by question type: point-in-time lookup, interval overlap, trend detection, and contradiction resolution.
• Capture provenance for each edge update so you can explain why a fact changed and which source triggered that change.
• Treat decay as a ranking layer, not deletion policy, unless regulatory requirements mandate hard retention limits.
• Simulate late-arriving events and backfills because production streams rarely arrive in perfect chronological order.

60-Day Delivery Plan

1. Week 1-2: define temporal schema and versioning policy, then generate synthetic event streams with backdated corrections.
2. Week 3-4: implement temporal retrieval API endpoints and validate answer consistency across equivalent query formulations.
3. Week 5-6: add staleness dashboards (invalidated edge rate, contradiction count, temporal miss rate) and incident runbooks.
4. Week 7-8: roll out to one agent cohort that depends on timeline reasoning and compare decision quality vs control group.

Failure Modes To Monitor

• Incorrect timestamp semantics creating false confidence in historical answers.
• Aggressive decay suppressing still-relevant facts and hurting recall on long-horizon queries.
• Lack of provenance making audits impossible for regulated workflows.
• Write amplification from frequent edge updates without compaction strategy.

Weekly Scoreboard

• Retrieval quality: Recall@k, answer faithfulness, and memory-hit attribution by workflow.
• Operational reliability: p95 retrieval latency, timeout rate, and failed consolidation jobs.
• Governance quality: policy-violation count, approval escalations, and unresolved audit findings.
• Business impact: task completion time, correction rate, and analyst intervention volume.