Structured memory representations such as knowledge graphs are central to autonomous agents and other long-lived systems. However, most existing approaches model time as discrete metadata, either sorting by recency (burying old-yet-permanent knowledge), simply overwriting outdated facts, or requiring an expensive LLM call at every ingestion step, leaving them unable to distinguish persistent facts from evolving ones. To address this, we introduce RoMem, a drop-in temporal knowledge graph module for structured memory systems, applicable to agentic memory and beyond. A pretrained Semantic Speed Gate maps each relation's text embedding to a volatility score, learning from data that evolving relations (e.g., "president of") should rotate fast while persistent ones (e.g., "born in") should remain stable. Combined with continuous phase rotation, this enables geometric shadowing: obsolete facts are rotated out of phase in complex vector space, so temporally correct facts naturally outrank contradictions without deletion. On temporal knowledge graph completion, RoMem achieves state-of-the-art results on ICEWS05-15 (72.6 MRR). Applied to agentic memory, it delivers 2-3x MRR and answer accuracy on temporal reasoning (MultiTQ), dominates hybrid benchmark (LoCoMo), preserves static memory with zero degradation (DMR-MSC), and generalises zero-shot to unseen financial domains (FinTMMBench).
Figures & tables
Figure 1: Performance overview. RoMem adds temporal reranking to the same HippoRAG graph-construction and semantic-retrieval pipeline. It improves temporal and hybrid retrieval while preserving performance on largely non-temporal memory. Table 2 reports results for both the closed- and open-source configurations.
Figure 2: Overview of the RoMem Architecture. The framework consists of four stages: (A) Functional Rotation (§ 3.3 ) applies geometric phase shifts to obsolete facts; (B) Semantic Speed Gate (§ 3.4 ) determines relational volatility from text embeddings; (C) Two-Phase Training (§ 3.5 ) pretrains the gate and learns the temporal spectrum; (D) Inference-Time Retrieval (§ 3.6 ) resolves contradictions via geometric shadowing.
Method
MRR
Hit@1
Hit@3
Hit@10
Non-Rotation Based
DistMult ( 2015 )
45.6
33.7
-
69.1
DE-SimplE ( 2020 )
51.3
39.2
57.8
74.8
TComplEx ( 2020 )
66.5
58.3
71.6
81.1
TLT-KGE ( 2022 )
68.6
60.7
73.5
83.1
HGE ( 2024a )
68.8
60.8
74.0
83.5
Table 1: Results on ICEWS05-15. Baseline results are taken from Li et al. (2025) and Shen et al. (2025) . Best results are in bold . Green cells indicate results where RoMem improves over its backbones (DistMult and ChronoR).
Table 2: Comprehensive evaluation of RoMem . (a) MultiTQ: temporal reasoning. (b) LoCoMo: hybrid retrieval (Recall@10; answer F1 is in Appendix F ). (c) DMR-MSC: largely non-temporal memory. (d) FinTMMBench: transfer to an unseen domain. Implementation denotes the graph-construction LLM and embedding model. Best results are in bold . Green cells mark the direct comparison with HippoRAG under the same graph-construction and semantic-retrieval pipeline.
Variant
MRR
H@3
Acc@5
Acc@10
Full RoMem
0.3371
0.384
0.366
0.392
w/o time contrastive
0.3254
0.368
0.340
0.390
w/o semantic gate
0.1044
0.116
0.100
0.108
Discrete time
0.0452
0.046
0.040
0.040
Table 3: Single-factor ablation on the fixed MultiTQ subset (selected metrics; full results in Appendix F.1 ). “No gate” uses one speed for all relations; “discrete time” uses learned timestamp lookups.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
RQ
Facts/Docs
Queries
Task Type
ICEWS05-15
RQ1
461,329
46,092
TKG completion
MultiTQ
RQ2
11,074
500
Temporal KGQA
LoCoMo
RQ2
10 conv.
1,986
Conv. memory QA
DMR-MSC
RQ2
500 dial.
500
Dialogue memory QA
FinTMMBench
RQ3
908 docs †
100 †
Financial temporal QA
Appendix
Table 4: Summary statistics of evaluation datasets.
Category
Full test set
Fixed subset
equal
17,311 (31.71%)
155 (31.00%)
before_after
11,073 (20.29%)
102 (20.40%)
first_last
10,480 (19.20%)
108 (21.60%)
equal_multi
3,207 (5.88%)
23 (4.60%)
after_first
6,266 (11.48%)
49 (9.80%)
before_last
6,247 (11.44%)
63 (12.60%)
Appendix
Table 5: Distribution of the full MultiTQ test set and the fixed 500-question evaluation subset.
Variant
MRR
H@1
H@3
H@10
Acc@5
Acc@10
Full RoMem
0.3371
0.252
0.384
0.502
0.366
0.392
w/o time contrastive
0.3254
0.238
0.368
0.504
0.340
0.390
w/o semantic gate
0.1044
0.084
0.116
0.142
0.100
0.108
Discrete time
0.0452
0.042
0.046
0.048
0.040
0.040
Appendix
Table 6: Full single-factor ablation on MultiTQ. “No gate” uses one temporal speed for every relation; “discrete time” replaces functional time with learned timestamp lookups.
Method
OpenAI F1
Server F1
Zep
0.343
0.320
Mem0
0.457
0.444
A-Mem
0.459
0.382
LicoMemory
0.346
0.306
HippoRAG
0.457
0.446
RoMem
0.457
0.447
Appendix
Table 7: Aggregate LoCoMo partial-match F1. “OpenAI” and “Server” refer to the graph-construction and embedding configurations; the answer reader is fixed.
Method
DMR-MSC original
DMR-MSC strict
FinTMMBench original
FinTMMBench strict
Zep
0.302
0.278
0.480
0.470
Mem0
0.858
0.834
0.550
0.540
A-Mem
0.848
0.830
0.540
0.540
LicoMemory
0.670
0.604
0.480
0.480
HippoRAG
0.852
0.832
0.550
0.550
RoMem
0.862
0.850
0.580
0.560
Appendix
Table 8: Acc@5 under the original and stricter conflict-aware answer judges. All saved answers are identical across judge protocols.
Parameter
Value
Description
temporal_backbone
chronor
ChronoR rotation backbone
chronor_k
3
Number of rotation sub-spaces
gamma
200.0
Embedding range: (γ+2)/dim
adversarial_temperature
1.0
Self-adversarial sampling temperature
regularization_weight
10−5
N3 per-batch regularization (ChronoR)
steps_per_update
500
Training epochs per update cycle
Appendix
Table 9: TKGE hyperparameters. Parameters above the mid-rule are core KGE settings; below are temporal extension parameters.
Figure 3: Scores for the competing slot (Obama, Consult, ?) . The light curves are the raw quarterly scores used for retrieval, and the bold curves are 5-quarter rolling averages shown only to make the overall trend easier to see. The blue and yellow regions mark the observation periods for Blair (2007-2008) and Xi (2013-2015), respectively. The red dot marks where the smoothed curves cross. The gate assigns a high rotation speed to “Consult” ( αr=0.87 ).
Relation
αr
Category
Seen during gate pretraining (ICEWS05-15)
Consult
0.87
Dynamic
Host a visit
0.86
Dynamic
Engage in negotiation
0.63
Dynamic
Sign formal agreement
0.53
Dynamic
Cooperate economically
0.16
Low
Appendix
Table 10: Pretrained semantic speed gate values. Higher αr = faster rotation. Seen : appeared in ICEWS05-15 during pretraining; Unseen : zero-shot via text embedding similarity.