arxiv: 2604.11544 · v1 · submitted 2026-04-13 · 💻 cs.CL · cs.AI

Recognition: unknown

Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs and Agentic Memory

Weixian Waylon Li , Jiaxin Zhang , Xianan Jim Yang , Tiejun Ma , Yiwen Guo

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:09 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords temporal knowledge graphsphase rotationagentic memorysemantic speed gategeometric shadowingknowledge graph completionstructured memorycontinuous time modeling

0 comments

The pith

Continuous phase rotation in complex vector space lets knowledge graphs update evolving facts while keeping permanent ones intact without deletion or overwriting.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing approaches to temporal knowledge in graphs either sort by recency and bury old facts, overwrite them outright, or call a large language model for every change. The paper proposes RoMem, which first uses a pretrained Semantic Speed Gate to turn each relation's text embedding into a volatility score, then rotates the corresponding complex vectors at speeds that match how fast each relation actually changes. Fast-rotating facts drift out of phase over time while slow ones stay aligned, so the system naturally ranks current correct statements above contradictions through geometry alone. This matters for long-lived autonomous agents because it preserves stable knowledge across many updates and avoids the cost of repeated language model interventions. The method is presented as a drop-in module that works for both standard temporal knowledge graph completion and agentic memory systems.

Core claim

By mapping relations to volatility scores via a pretrained Semantic Speed Gate and then applying continuous phase rotation in complex space, obsolete facts rotate out of phase and are geometrically shadowed so that temporally correct facts outrank contradictions without any explicit deletion or overwriting.

What carries the argument

The Semantic Speed Gate that converts relation text embeddings into volatility scores, which then control the speed of continuous phase rotation in complex vector space to produce geometric shadowing.

Load-bearing premise

The pretrained Semantic Speed Gate correctly distinguishes fast-evolving relations from stable ones based on text embeddings, and the resulting differential phase rotations create geometric shadowing that reliably ranks temporally correct facts highest.

What would settle it

Run the system on a temporal dataset that contains known contradictions between persistent and rapidly changing facts, but replace the learned volatility scores with random values, and check whether the ranking of correct current facts collapses.

Figures

Figures reproduced from arXiv: 2604.11544 by Jiaxin Zhang, Tiejun Ma, Weixian Waylon Li, Xianan Jim Yang, Yiwen Guo.

**Figure 1.** Figure 1: Performance Overview. Our main contributions are as follows: • We formalise the static-dynamic dilemma in graph-based agentic memory, showing that discrete timestamp metadata treats all relations identically, preventing temporal conflict resolution without sacrificing static knowledge. • We formulate temporal conflict resolution as continuous geometric shadowing in complex vector space, replacing destruc… view at source ↗

**Figure 2.** 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.… view at source ↗

**Figure 3.** Figure 3: Scoring trace for the competing slot (Obama, Consult, ?). Bold curves are smoothed (5-quarter rolling average); light traces show raw quarterly scores. The blue and yellow shaded regions mark the observation windows for Blair (2007–2008) and Xi (2013–2015) respectively. The crossover point τ ∗ (red dot) marks where the temporally outdated fact is geometrically shadowed by the newer one. The gate value αr =… view at source ↗

read the original abstract

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).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RoMem's phase-rotation idea for temporal KGs is a clean geometric shift from deletion or recency, but the mechanism lacks visible derivation or ablations to confirm it actually ranks facts correctly.

read the letter

The paper's core move is to treat time as continuous phase rotation in complex space rather than discrete labels or overwrites. A Semantic Speed Gate maps relation text embeddings to per-relation volatility scores, then each stored fact embedding gets rotated by speed times elapsed time. The claim is that this produces geometric shadowing so that, at query time, the right fact naturally scores higher than contradictions without any explicit deletion step. That framing is new for agentic memory modules and temporal KGs; it directly targets the stable-versus-evolving knowledge problem in long-lived systems without constant LLM calls or recency bias burying old facts. The reported numbers give it some weight: SOTA 72.6 MRR on ICEWS05-15, 2-3x gains on MultiTQ temporal reasoning, strong results on LoCoMo and DMR-MSC with zero static-memory degradation, and zero-shot transfer to FinTMMBench. Those benchmarks cover both standard TKG completion and practical agent settings, which is useful. The approach is presented as a drop-in module, which lowers the barrier for adoption. The main soft spot is the mechanism itself. No equation, toy example, or proof sketch shows how the per-fact phase offset interacts with the query embedding to produce the desired ranking; the abstract and stress-test note both leave this asserted rather than derived. The volatility gate is learned from data, so its reliability across domains and entity pairs is central but not obviously stress-tested with ablations or error analysis in the visible material. If the full paper supplies the missing math and controls, that gap closes; otherwise the performance claims rest on an unverified geometric effect. This is for people building structured memory in autonomous agents or working on temporal knowledge representation. A reader focused on practical long-term memory systems would find the benchmarks and the continuous-time framing worth examining. It deserves peer review because the problem is real, the geometric angle is distinct from prior discrete or recency methods, and the empirical scope is broad enough to justify referee time even if the technical grounding needs tightening.

Referee Report

3 major / 2 minor

Summary. The paper introduces RoMem, a drop-in temporal knowledge graph module for agentic memory and structured systems. It uses a pretrained Semantic Speed Gate to map each relation's text embedding to a volatility score (fast for evolving relations like 'president of', slow for persistent ones like 'born in'), then applies continuous phase rotation in complex vector space so that obsolete facts rotate out of phase and are geometrically shadowed. This allows temporally correct facts to outrank contradictions at query time without deletion, overwriting, or per-ingestion LLM calls. The approach is claimed to achieve SOTA 72.6 MRR on ICEWS05-15 temporal KG completion and 2-3x gains in MRR/accuracy on MultiTQ temporal reasoning, plus strong results on LoCoMo, DMR-MSC (zero static degradation), and zero-shot FinTMMBench.

Significance. If the geometric mechanism is shown to work as described, the result would be significant for long-lived autonomous agents: it offers a parameter-light, deletion-free way to maintain evolving structured memory that avoids both recency bias and expensive per-step LLM reasoning, with potential for broad applicability beyond the reported benchmarks.

major comments (3)

[Methods / geometric shadowing description] The central claim that continuous phase rotation produces geometric shadowing (obsolete facts automatically outranked by correct ones via dot-product similarity) is load-bearing but unsupported by derivation. No equation, toy example, or proof sketch is provided showing how the per-fact phase offset (volatility score multiplied by time since insertion) interacts with a query embedding to guarantee the desired ranking without explicit time labels or deletion.
[Experiments / results tables] Performance numbers (72.6 MRR on ICEWS05-15, 2-3x MRR/accuracy on MultiTQ, dominance on LoCoMo) are stated without derivation details, training procedure for the Semantic Speed Gate, baseline comparisons, ablation studies, or error analysis. This makes it impossible to verify whether reported gains are attributable to the proposed rotation mechanism rather than implementation specifics or data artifacts.
[Semantic Speed Gate / §3.1] The Semantic Speed Gate learns volatility scores from data, creating a dependence on the training distribution that contradicts the claim of a purely geometric, time-label-free solution. The paper must clarify whether the gate is fixed after pretraining on an independent corpus or fitted to the target datasets, and provide cross-domain ablations.

minor comments (2)

[Notation / preliminaries] Notation for complex embeddings and phase multiplication should be made explicit (e.g., define the rotation operator and similarity function used at query time).
[Introduction / abstract] The abstract and introduction assert 'parameter-free' geometric behavior, but the volatility scores are learned parameters; this tension should be resolved in the text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important areas for clarification and expansion, particularly around the geometric mechanism and experimental transparency. We address each major comment point by point below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Methods / geometric shadowing description] The central claim that continuous phase rotation produces geometric shadowing (obsolete facts automatically outranked by correct ones via dot-product similarity) is load-bearing but unsupported by derivation. No equation, toy example, or proof sketch is provided showing how the per-fact phase offset (volatility score multiplied by time since insertion) interacts with a query embedding to guarantee the desired ranking without explicit time labels or deletion.

Authors: We agree that the manuscript would benefit from an explicit derivation of the geometric shadowing effect to make the central claim fully rigorous. In the revised version, we will add a new subsection in the Methods section containing: (i) the core equation where a fact embedding f with volatility v is rotated to f * exp(i * v * Δt) and similarity is computed as Re(q · rotated_f); (ii) a short proof sketch demonstrating that for sufficiently large Δt and v > 0 the real-part dot product falls below that of the temporally correct fact; and (iii) a minimal toy example with two contradictory facts differing only in insertion time. These additions will directly show how the phase offset produces the desired ranking without requiring time labels at query time. revision: yes
Referee: [Experiments / results tables] Performance numbers (72.6 MRR on ICEWS05-15, 2-3x MRR/accuracy on MultiTQ, dominance on LoCoMo) are stated without derivation details, training procedure for the Semantic Speed Gate, baseline comparisons, ablation studies, or error analysis. This makes it impossible to verify whether reported gains are attributable to the proposed rotation mechanism rather than implementation specifics or data artifacts.

Authors: We acknowledge that the current experimental section lacks sufficient detail for independent verification. In the revision we will expand the Experiments section to include: the full training procedure and hyperparameters for the Semantic Speed Gate (pretrained on an independent corpus), complete baseline tables with all compared methods, ablation studies that isolate the contribution of continuous phase rotation (e.g., setting v = 0), and an error analysis breaking down failure modes on ICEWS05-15. These additions will allow readers to confirm that the reported gains arise from the geometric mechanism. revision: yes
Referee: [Semantic Speed Gate / §3.1] The Semantic Speed Gate learns volatility scores from data, creating a dependence on the training distribution that contradicts the claim of a purely geometric, time-label-free solution. The paper must clarify whether the gate is fixed after pretraining on an independent corpus or fitted to the target datasets, and provide cross-domain ablations.

Authors: The Semantic Speed Gate is pretrained once on an independent corpus and kept fixed at inference time on all target datasets; this is stated in §3.1 and ensures that the core operation remains purely geometric and time-label-free after the one-time pretraining step. Volatility is derived from relation semantics rather than being refitted per benchmark. To address the referee's request, the revised manuscript will add cross-domain ablation results showing performance when the gate (pretrained on general text) is applied zero-shot to ICEWS05-15, MultiTQ, and FinTMMBench without any target-domain fine-tuning. revision: yes

Circularity Check

0 steps flagged

No significant circularity; mechanism and results are empirically grounded rather than self-referential

full rationale

The paper presents RoMem as an empirical architecture: a pretrained Semantic Speed Gate learns per-relation volatility scalars from training data, which are then used to drive continuous phase rotation in complex embeddings for geometric shadowing. Performance is reported via standard held-out evaluation on public benchmarks (ICEWS05-15, MultiTQ, LoCoMo, DMR-MSC, FinTMMBench). No derivation chain is offered in which a claimed prediction or first-principles result is shown to equal its own inputs by construction; the volatility mapping is a fitted component whose outputs are tested on disjoint data, and the rotation step is an independent geometric operation. This is ordinary supervised modeling plus a novel inductive bias, not a closed loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on a learned mapping from embeddings to volatility and on the assumption that phase rotation produces useful shadowing; both are introduced without external benchmarks in the abstract.

free parameters (1)

volatility scores per relation
Produced by the Semantic Speed Gate pretrained on data; each relation receives a fitted speed that controls its rotation rate.

axioms (1)

domain assumption Complex vector space permits continuous phase rotation to represent temporal evolution of facts
Invoked to enable geometric shadowing of obsolete facts.

invented entities (1)

Semantic Speed Gate no independent evidence
purpose: Maps relation text embedding to a volatility score that sets rotation speed
New component introduced to learn which relations evolve quickly versus slowly.

pith-pipeline@v0.9.0 · 5557 in / 1503 out tokens · 78189 ms · 2026-05-10T16:09:31.785467+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 3 canonical work pages · 1 internal anchor

[1]

InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 13870, Bangkok, Thailand

Evaluating very long-term conversational memory of LLM agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 13870, Bangkok, Thailand. Association for Compu- tational Linguistics. Meta AI. 2024. Introducing meta llama 3: The most capable openly available llm to date. Charles...

2024
[2]

MemGPT: Towards LLMs as Operating Systems

Memgpt: Towards llms as operating systems. Preprint, arXiv:2310.08560. Jiaxin Pan, Mojtaba Nayyeri, Yinan Li, and Steffen Staab. 2024a. Hge: Embedding temporal knowledge graphs in a product space of heterogeneous geometric subspaces.Proceedings of the AAAI Conference on Artificial Intelligence, 38(8):8913–8920. Shirui Pan, Linhao Luo, Yufei Wang, Chen Che...

work page internal anchor Pith review Pith/arXiv arXiv 2024
[3]

Memotime: Memory-augmented temporal knowledge graph enhanced large language model reasoning.arXiv preprint arXiv:2510.13614, 2025a

MemInsight: Autonomous memory augmenta- tion for LLM agents. InProceedings of the 2025 Con- ference on Empirical Methods in Natural Language Processing, pages 33124–33140, Suzhou, China. As- sociation for Computational Linguistics. Jinqing Shen, Chengjin Xu, Yingqi Liu, Xuhui Jiang, Jiaming Li, Zhenxin Huang, Jens Lehmann, and Xuesong Chen. 2025. Learning...

work page arXiv 2025
[4]

Agent learning via early experience.arXiv preprint arXiv:2510.08558, 2025

Agentfold: Long-horizon web agents with proactive context management. Rui Ying, Mengting Hu, Jianfeng Wu, Yalan Xie, Xi- aoyi Liu, Zhunheng Wang, Ming Jiang, Hang Gao, Linlin Zhang, and Renhong Cheng. 2024. Simple but effective compound geometric operations for tem- poral knowledge graph completion. InProceedings of the 62nd Annual Meeting of the Associat...

work page arXiv 2024
[5]

Exact match: normalised entity string equal- ity
[6]

Containment: bidirectional substring check (gold⊆prediction or prediction⊆gold)
[7]

Advanced normalisation: strip prefixes (e.g., “The”), brackets, and punctuation, then sub- string match
[8]

January 2013

Time format matching: year–month level matching with month-name support (e.g., “January 2013”≈“2013-01”)

2013
[9]

Multi-answer: comma-separated answer parts are matched individually
[10]

Semantic overlap: word overlap >50% be- tween prediction and gold answer tokens
[11]

E Implementation Configurations and Hyperparameters E.1 Configurations We evaluate all agentic memory benchmarks under two implementation configurations:

Loose match: remove all spaces and under- scores, then substring match. E Implementation Configurations and Hyperparameters E.1 Configurations We evaluate all agentic memory benchmarks under two implementation configurations:
[12]

This configuration tests per- formance with API-based models

OpenAI: GPT-5-mini for graph construction (NER + OpenIE) and text-embedding-3-small for embedding. This configuration tests per- formance with API-based models
[13]

Server: LLaMA-3.1-70B-Instruct (Meta AI,
[14]

This configuration demonstrates open-source reproducibility

served via vLLM for graph construction and BAAI/BGE-M3 (Chen et al., 2024) for embedding. This configuration demonstrates open-source reproducibility. 16 In both configurations, theanswer LLMand LLM judgealways use GPT-5.2 via the OpenAI API to ensure fair comparison across all baselines. Baseline systems.We compare against: • Mem0(Chhikara et al., 2025):...

2024
[15]

born in” or “species

As τq moves forward, Blair’s score initially dominates but progressively decreases as the phase difference |τq −t Blair| grows. Simultaneously, Xi’s score rises as τq approaches his observed period. The crossover occurs around 2009, after which Xi geometricallyshadowsBlair. Note that the crossover point τ ∗ is not necessarily at the mid- point of the two ...

2009