Verbal memory retrieval carries rich diagnostic information for Alzheimer's disease through the order, hesitation, and categorical organization of recalled words, yet computational modeling of this retrieval process remains largely unexplored. A key challenge is that timing and hierarchical structure are diagnostically entangled: the diagnostic meaning of retrieval timing depends on its hierarchical context. We propose TrajLift, a spectral framework that models verbal recall as timed trajectories on a hierarchical graph. Our central mechanism is latency-modulated heat diffusion on the semantic hierarchy, a unified operator that lifts each recalled word into a representation of both multi-scale hierarchical structure and continuous retrieval timing. Formal analysis shows that the resulting representation is first-order separable into structural and temporal components with spectral selectivity across hierarchical levels. On synthetic benchmarks requiring structure-time interaction, TrajLift stays near-perfect on deep hierarchies where all baselines fail. On real-world clinical data, we show that TrajLift outperforms diverse baselines by jointly modeling timing and semantic structure, with gains that grow as retrieval timing becomes more informative.
Figures & tables
Figure 1: Verbal memory retrieval depicted as a timed trajectory on a three-level semantic hierarchy. Arrows indicate recall order; Δti denotes the latency between consecutive items.
Paradigm
Representative
Static Graph
Hierarchy
Continuous Time
Time-Aware Sequence
Time-LSTM [ 8 ]
-
✗
✓
Dynamic Graph Learning
TGN [ 9 ] , TGAT [ 10 ]
✗
✗
✓
Session/Trajectory GNN
SR-GNN [ 11 ] , TrajGAT [ 12 ]
✓
✗
✗
Hierarchical Pooling
DiffPool [ 13 ] , SAGPool [ 14 ]
✓
✓ (learned)
✗
Ours
Proposed
✓
✓ (explicit)
✓
Table 1: Inductive biases of existing approaches for trajectory learning on graphs.
Variant
Levels
Structure
Cluster Configuration
ST-XOR
2
16→64
16 clusters
ST-XOR-w
2
32→256
32 clusters ( w ider)
ST-XOR-wh
3
4→32→256
4 super- / 32 sub- ( w ider + h igher)
Table 2: Structural specifications of the hierarchy variants in the ST-XOR benchmark.
Benchmark
Time-informed Sequential Models
Dynamic Graph Models
Time-informed Graph Models
Time-informed Graph +Sequential Models
TrajLift (Ours)
LSTM +
Transformer +
TGN
TGAT
TrajGAT +
Graphormer +
GAT + w/ LSTM
R-GCN + w/ LSTM
ST-XOR
67.78 ± 0.43
54.22 ± 4.13
75.56 ± 5.35
99.78 ± 0.16
95.89 ± 1.03
73.56 ± 2.11
98.78 ± 0.42
99.44 ± 0.31
98.00 ± 0.98
ST-XOR-w
52.56 ± 0.96
50.11 ± 2.74
51.67 ± 2.13
91.56 ± 1.10
82.78 ± 1.34
55.00 ± 3.34
99.33 ± 0.51
98.78 ± 0.96
99.22 ± 0.85
ST-XOR-wh
49.78 ± 0.31
49.44 ± 2.38
49.33 ± 2.99
47.44 ± 2.08
49.56 ± 0.68
50.67 ± 1.36
48.89 ± 1.57
49.56 ± 0.68
98.67 ± 0.57
Table 3: Classification accuracy (%) on the ST-XOR synthetic benchmark, mean ± SD over five random seeds. Accuracies above 90% are underlined.
HKLLT-4
HKLLT-5
Model
Acc
F1
Sens.
Acc
F1
Sens.
NCD Detection Baselines
Recall count (LR)
74.96 ± 4.59
74.02 ± 4.36
73.55 ± 11.71
75.90 ± 4.39
75.19 ± 4.53
74.84 ± 7.34
Low-level features (LR, 7-D)
74.79 ± 4.07
73.99 ± 4.28
73.12 ± 8.93
75.90 ± 3.97
75.07 ± 4.00
73.33 ± 7.26
BERT
72.05 ± 5.10
70.00 ± 5.20
58.06 ± 6.90
74.19 ± 3.94
72.33 ± 4.32
61.29 ± 7.42
RoBERTa
75.56 ± 4.56
73.51 ± 5.17
60.97 ± 9.35
73.76 ± 3.77
71.74 ± 4.30
60.11 ± 9.03
Table 4: NC-vs-NCD classification on HKLLT delayed recall (%), mean ± SD over 15 splits. Paired comparisons are shown in Figure 2 . Sens. = sensitivity. Best result per column in bold.
Figure 2: Paired differences between TrajLift and selected strong baselines, with 95% confidence intervals. Wins: splits on which TrajLift is better.
Input latencies
HKLLT-4
HKLLT-5
Original
77.01 ± 5.68
80.85 ± 4.73
δk→0
73.08 ± 4.05 ( − 3.93)
74.02 ± 2.35 ( − 6.84)
δk→δˉ
69.83 ± 6.94 ( − 7.18)
63.59 ± 5.03 ( − 17.26)
Table 5: Test-time latency perturbation (accuracy, %), mean ± SD over an independent set of 15 splits. Parentheses: change from the unperturbed input.
Figure 3: (a) Heat kernel filter response at three scales. (b) Diffused representations preserve the ultrametric distance in cosine similarity. (c) Temporal modulation effect: embedding distance increases with latency. (d) Time order sensitivity: prediction divergence grows with shuffle ratio. All panels use a TrajLift model trained on one HKLLT-4 split.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
CV of cluster size
0
0.232
0.524
0.970
TrajLift accuracy (%)
98.67 ± 0.57
98.67 ± 1.18
98.87 ± 0.61
99.00 ± 0.41
Appendix
Table 6: Unbalanced hierarchies on ST-XOR-wh. CV: coefficient of variation of cluster sizes.
σ
0
0.25
0.50
0.75
1.00
1.50
2.00
d′
9.77
5.79
3.37
2.33
1.77
1.19
0.90
Overlap (%)
0
0.4
9.2
24.4
37.7
55.2
65.4
TrajLift accuracy (%)
98.67
99.13
97.87
95.00
79.40
60.80
53.53
Appendix
Table 7: Test-time timestamp perturbation on ST-XOR-wh.
Scales
Number
Accuracy
{macro}
1
99.67 ± 0.75
{meso, macro}
2
99.47 ± 0.30
{micro, macro}
2
99.27 ± 0.60
{micro, meso, macro}
3
98.67 ± 0.57
{meso}
1
94.47 ± 3.74
{micro, meso}
2
53.20 ± 2.33
Appendix
Table 8: TrajLift with subsets of diffusion scales on ST-XOR-wh (accuracy, %), mean ± SD over five random seeds.
Domain
Category
Items
Abstract
Relatives
Grandmother, Aunt, Niece, Cousin
Countries
India, Switzerland, Laos, Chile
Concrete
Vegetables
Eggplant, Cucumber, Mustard Greens, Onion
Furniture
Desk, Mirror, Lamp, Wardrobe
Appendix
Table 9: HKLLT word list and semantic hierarchy. The 16 target words are organized into four categories under two domains.
NC
NCD
Number
233
154
Age (years)
70.8 ± 6.5
74.5 ± 7.8
Education (years)
8.8 ± 4.1
7.2 ± 4.4
Gender (M/F)
98/135
74/80
Appendix
Table 10: Demographic characteristics of participants. Continuous variables are reported as mean ± SD; gender as Male/Female counts.
Measure
Trial
NC
NCD
t
p
Semantic Clustering
T4
3.33±2.99
1.38±2.23
6.91
<.001
T5
2.99±2.61
1.02±1.47
8.48
<.001
Recall Count
T4
6.82±2.50
3.21±2.50
13.87
<.001
T5
6.36±2.62
2.69±2.25
14.18
<.001
Recency Effect
T4
0.12±0.12
0.12±0.20
−0.02
.984
T5
0.12±0.11
0.13±0.22
−0.65
.517
Appendix
Table 11: Measure validation via group comparison ( t -test). Values are mean ± SD.
Statistic
Trial
NC
NCD
Sequence Length
T4
9.06±4.60
4.52±4.24
T5
7.76±3.86
3.42±3.06
Mean Inter-Item Latency (s)
T4
2.70±1.38
3.27±2.14
T5
2.39±1.25
3.49±2.08
Word Coverage
T4
6.82±2.50
3.21±2.50
T5
6.36±2.62
2.69±2.25
Appendix
Table 12: Trajectory statistics by trial and group. Values are mean ± SD.
Figure 4: Distribution of inter-item latencies in the HKLLT dataset. (a,c) Raw latencies δk for Trials 4 and 5: the distributions are heavily right-skewed (mean ≫ median) with 99th percentiles exceeding 30s. (b,d) Log-transformed latencies ln(δk) : the distributions become approximately symmetric. Latencies exceeding 20s are clipped before transformation.
Figure 5: Example recall trajectories from HKLLT Trial 4 (Class 1 = NCD, Class 0 = NC). Words are laid out by category. Arrows follow recall order and are labeled with the inter-item latency in seconds.
Figure 6: Example recall trajectories from HKLLT Trial 5 (Class 1 = NCD, Class 0 = NC). Words are laid out by category. Arrows follow recall order and are labeled with the inter-item latency in seconds.
Large Language Models (LLMs) suffer from temporal misalignment, often due to the contradictory nature of their training corpora. While current mitigation strategies rely on computationally expensive fine-tuning or context-heavy retrieval augmented generation (RAG), the internal mechanisms governing time-sensitive recall remain under-explored. Unlike prior studies that identify temporal components such as attention heads and MLP layers, we provide the first feature-level map of temporal recall by isolating individual MLP features via transcoder circuit tracing. We identify three node categories (common temporal, common to the year, and chrono-semantic) which interact to generate a temporal filter during factual recall. By analysing Gemma 2 2B, LLaMA 3.2 1B, and Qwen3-4B, we show that these features do not follow a simple linear pipeline but represent time through a parallel and mixed syntactic-semantic interplay across layers. We additionally discover a class of higher-layer temporal components invisible to existing EAP-IG methods, establishing transcoders as a more complete lens for temporal interpretability in time-sensitive factual recall. These findings present MLP components for potential targeted interventions in time-sensitive factual recall
AI Memory, specifically how models organizes and retrieves historical messages, becomes increasingly valuable to Large Language Models (LLMs), yet existing methods (RAG and Graph-RAG) primarily retrieve memory through similarity-based mechanisms. While efficient, such System-1-style retrieval struggles with scenarios that require global reasoning or comprehensive coverage of all relevant information. In this work, We propose Mnemis, a novel memory framework that integrates System-1 similarity search with a complementary System-2 mechanism, termed Global Selection. Mnemis organizes memory into a base graph for similarity retrieval and a hierarchical graph that enables top-down, deliberate traversal over semantic hierarchies. By combining the complementary strength from both retrieval routes, Mnemis retrieves memory items that are both semantically and structurally relevant. Mnemis achieves state-of-the-art performance across all compared methods on long-term memory benchmarks, scoring 93.9 on LoCoMo and 91.6 on LongMemEval-S using GPT-4.1-mini.
Retrieval-augmented language model systems generally select context by query-conditioned relevance. We argue that long-lived AI systems require a second operator: activation-conditioned memory availability. We specify an architecture in which request and goal cues seed an authorized memory graph, activation and fan attenuation allocate a bounded traversal frontier, retrieved entities recursively emit cues, and a typed evidence graph is reconstructed before generation. The specification yields five architectural consequences about candidate selection, traversal control, graph regularization, evidence reconstruction, and reinforcement. We derive implementation-level cost estimates and define ablations, failure conditions, and a reproducible evaluation protocol. The central claim is narrowed or falsified if a simpler graph reranker matches full traversal on causal and temporal recovery under equal budgets. This architecture-and-analysis preprint reports no benchmark measurements. If the claim holds, retrieval returns the evidence path that makes a memory available, not only a ranked list.