TAAM:Inductive Graph-Class Incremental Learning with Task-Aware Adaptive Modulation DOI: WSBG8309 Conference: Proc. of the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026); May 25 – 29, 2026; Paphos, Cyprus 1797 CCS: Computing methodologies Knowledge representation and reasoning CCS: Computing methodologies Neural networks CCS: Computing methodologies Multi-agent planning CCS: Information systems Data mining Jingtao Liu, Xinming Zhang Abstract . Graph Continual Learning (GCL) aims to solve the challenges of streaming graph data. However, current methods often depend on replay-based strategies, which raise concerns like memory limits and privacy issues, while also struggling to resolve the stability-plasticity dilemma. In this paper, we suggest that lightweight, task-specific modules can effectively guide the reasoning process of a fixed GNN backbone. Based on this idea, we propose T ask- A ware A daptive M odulation (TAAM). The key component of TAAM is its lightweight Neural Synapse Modulators (NSMs). For each new task, a dedicated NSM is trained and then frozen, acting as an “expert module.” These modules perform detailed, node-attentive adaptive modulation on the computational flow of a shared GNN backbone. This setup ensures that new knowledge is kept within compact, task-specific modules, naturally preventing catastrophic forgetting without using any data replay. Additionally, to address the important challenge of unknown task IDs in real-world scenarios, we propose and theoretically prove a novel method named Anchored Multi-hop Propagation (AMP). Notably, we find that existing GCL benchmarks have flaws that can cause data leakage and biased evaluations. Therefore, we conduct all experiments in a more rigorous inductive learning scenario. Extensive experiments show that TAAM comprehensively outperforms state-of-the-art methods across eight datasets.Code and Datasets are available at: https://github.com/1iuJT/TAAM_AAMAS2026 . C. Amato, L. Dennis, V. Mascardi, J. Thangarajah (eds.) 1. Introduction While Graph Neural Networks (GNNs) have demonstrated remarkable success on static graphs, real-world graph data is often dynamic and evolving.To address this, Continual Graph Learning (CGL) has become a key research field ( 27 ) , where models must learn from an evolving stream of tasks while retaining previously acquired knowledge (Figure 1 , top). Among CGL paradigms, Graph Class-Incremental Learning (GCIL) presents a particularly challenging and practical frontier. Its core tenet is that task identifiers (IDs) are unavailable during inference, a scenario mirroring real-world applications like drug discovery, where new compound families emerge ( 34 ) , or social networks, where user communities evolve over time ( 28 ) . This task-agnostic requirement introduces a dual challenge: first, overcoming catastrophic forgetting, and second, resolving the “inter-task class separation” problem—the difficulty of distinguishing between similar classes from different tasks without explicit context ( 25 ) . Figure 1 . Top: A typical CGL scenario where a model must adapt to a stream of new tasks (indicated by new node colors) in an evolving graph. Bottom: An illustration of the Stability-Plasticity Dilemma on the Citeseer benchmark. While strong replay-based methods effectively mitigate the catastrophic forgetting of early tasks, they exhibit diminished plasticity. A two-part diagram. The top illustrates a continuous stream of graph data with new nodes and tasks. The bottom shows a bar chart comparing performance, highlighting the stability-plasticity dilemma where replay methods suffer from reduced plasticity. To address these challenges, researchers have primarily explored three paradigms: regularization-based, replay-based, and parameter-isolation methods ( 34 ; 32 ) . Replay-based approaches, which store and rehearse historical data, can effectively mitigate forgetting but introduce significant computational and storage overhead, along with potential data privacy risks. Critically, they face a fundamental trade-off between preserving old knowledge (stability) and acquiring new information (plasticity)(Figure 1 , Bottom).In contrast, parameter-isolation methods offer a more direct path to overcoming forgetting. By dedicating distinct parameters to different tasks, these approaches structurally prevent knowledge interference ( 32 ; 37 ) . Recently, Prompt Learning has emerged as a highly effective parameter-efficient fine-tuning strategy, achieving great success in computer vision ( 34 ; 32 ) . This paradigm was quickly adapted to the graph domain ( 17 ; 29 ) , with pioneering works like TPP conditioning a frozen GNN with small, learnable "graph prompts" based on inferred task IDs ( 20 ) .However, the effectiveness of this paradigm is often predicated on the availability of a powerful, pre-trained GNN backbone to ensure robust knowledge transferability ( 3 ; 11 ) . Moreover, many graph prompting methods employ universal prompts that offer only coarse-grained control over the GNN’s internal computational flow, thus limiting their adaptive capacity ( 20 ) . This reliance on pre-training presents unique challenges in the graph domain. Unlike in NLP and Computer vision standardized, universal pre-trained GNNs are not yet readily available, and aligning knowledge across heterogeneous graph structures remains a formidable task ( 7 ; 26 ) . This confluence of challenges motivates a fundamental research question: Is it possible to design a CGL approach that eliminates the dependency on a heavily pre-trained model, yet still achieves high performance in a efficient, replay-free manner? In this paper, we tackle this challenge by rethinking how the stability-plasticity trade-off can be managed in a resource-efficient, replay-free context. Our approach maintains stability by leveraging a frozen GNN backbone and encapsulating task-specific knowledge within lightweight modulators. We argue that the key to unlocking plasticity lies in designing highly expressive, task-specific modules and, critically, a robust mechanism to identify and deploy the correct knowledge for a given task at inference time. To this end, we introduce T ask- A ware A daptive M odulation (TAAM), a novel approach that operationalizes this principle. The architecture of TAAM is composed of two core innovations: First, for plasticity, we design dynamic and lightweight N eural S ynapse M odulators (NSMs). For each task, a new NSM is created and trained to perform fine-grained, node-attentive modulation on the internal representations of the frozen GNN backbone. These modules act as task-specific experts, effectively steering the GNN’s computational flow without altering its stable, foundational parameters. Second, for stability, we address the critical challenge of task-ID inference. We propose and theoretically ground a novel method named A nchored M ulti-hop P ropagation (AMP) to generate highly robust and discriminative task prototypes. For inference, AMP allows TAAM to accurately identify the task affiliation of incoming data, ensuring the correct expert NSM is selected. This precise retrieval mechanism is fundamental to preventing catastrophic forgetting. The modulated output is then processed by a unified classifier whose output dimension expands for new classes, preserving knowledge at the decision layer.We summarize our main contributions as follows. • We propose TAAM , a novel CGL approach that operates without reliance on data replay or extensive model pre-training, offering a truly resource-efficient solution. • We introduce and provide theoretical grounding for Anchored Multi-hop Propagation (AMP), a robust method for task-ID inference that ensures model stability and prevents forgetting, validated across eight datasets. • We propose a lightweight, and node-attentive modulator (NSM) that effectively incorporates task-specific knowledge and provides more expressive, fine-grained control over GNN representations than traditional static prompting methods. • We identify limitations in existing GCL benchmarks and are the first to validate our approach in the more challenging and realistic inductive graph continual learning scenario, demonstrating its practical effectiveness. 2. Related Works Replay-based Methods. A dominant paradigm in Continual Graph Learning (CGL) is the use of replay-based methods, which maintain a memory buffer of representative exemplars from past tasks. During training on a new task, these stored exemplars are rehearsed alongside the new data, mitigating catastrophic forgetting through joint optimization ( 15 ; 16 ; 36 ; 33 ) . The sophistication of what is stored has evolved, from representative nodes ( 39 ) to more complex structures. For instance, CaT ( 15 ) and PUMA ( 16 ) focus on graph condensation techniques to create small, synthetic graphs for replay. Others compress topological information, either by storing sparsified computation subgraphs (SSM ( 36 ) ) or by decoupling the model from the graph via compressed topology-aware embeddings (PDGNNs ( 33 ) ). DeLoMe ( 21 ) constructs a memory composed of lossless node prototypes, aiming to fully capture and store the graph information of old tasks, and further proposes a de-biased GCL loss function to address category imbalance. While these methods often demonstrate strong performance, their reliance on storing and re-processing historical data introduces significant computational and memory overhead, along with potential data privacy concerns. Prompt Learning and Parameter-Efficient Methods. An alternative direction is rooted in Parameter-Efficient Transfer Learning (PETL), a paradigm where large pre-trained models are adapted to new tasks by updating only a small subset of parameters. Representative methods like Adapters achieve this by inserting lightweight modules between the layers of a backbone network to reduce fine-tuning costs ( 6 ; 4 ) . Building on this foundation, prompt learning has emerged as a leading PETL technique. It guides a model’s behavior for specific tasks by injecting learnable prompts,” and is widely adopted for its excellent few-shot and transferability performance under thepre-train, prompt, predict” framework ( 14 ) . Inspired by its success, researchers have recently extended this paradigm to Graph Neural Networks (GNNs). For instance, UPT conditions a frozen GNN on structural prompts to enhance few-shot and cross-domain performance ( 3 ) , while GraphPrompt unifies pre-training and downstream tasks through a shared interface to improve generalization ( 17 ) . Other works like TPP infers task IDs to select the appropriate task-specific prompt for conditioning the GNN ( 20 ) . While these methods achieve impressive parameter-efficient knowledge transfer, their effectiveness critically depends on the availability of a powerful, often resource-intensive pre-trained backbone. This prerequisite poses a significant challenge in the graph domain, where universally effective and readily available pre-trained GNNs are not as common as in other fields like NLP or vision. This limitation motivates our research into a new paradigm that delivers the parameter efficiency of prompt-based conditioning, but without the dependency on a pre-trained model, thereby offering a more flexible and accessible solution for continual graph learning. 3. Preliminary Figure 2 . Overview of TAAM framework.TAAM is inserted the frozen GNN backbone,to steer a GNN’s reasoning process in stream of graph data.The diagram illustrates the strategy of TAAM.Training Phase: For a new task \mathcal{G}_{k} , a new Neural Synapse Modulator ( \mathrm{NSM}_{k} ) is created from scratch and trained. Concurrently, its corresponding task prototype P_{k} is generated and stored.After training is complete, \mathrm{NSM}_{k} frozen and added to the TAAM module bank.Inference Phase:For incoming test graph \mathcal{G}^{k}_{t} with an unknown task ID , a prototype P^{k}_{t} generated using AMP,for selects the most relevant “expert" NSM. A flowchart of the TAAM framework. It shows the training phase where a task-specific NSM and prototype are created, and the inference phase where the AMP method identifies the unknown task ID to select the correct expert module. 3.1. Problem Formulation The work of this paper is oriented towards Graph Class-Incremental Learning (GCIL), where a model must continually learn new classes of nodes from a sequence of tasks. Formally, we are given a sequence of N tasks, \mathcal{T}=\{\tau_{1},\tau_{2},\dots,\tau_{N}\} . Each task \tau_{k} introduces a new set of node classes \mathcal{C}_{k} , where all class sets are disjoint, i.e., \mathcal{C}_{i}\cap\mathcal{C}_{j}=\emptyset for any i\neq j . After learning up to task \tau_{k} , the model f must be able to classify nodes from the union of all observed classes, \mathcal{C}_{1:k}=\bigcup_{i=1}^{k}\mathcal{C}_{i} . A critical constraint in the GCIL setting is that the task identity of data is unavailable during inference. Unlike earlier GCL benchmarks ( 34 ) , whichoperate in a transductive scenario (where training and testing nodes reside within the same graph), our study tackles the more demanding and practical inductive setting. Here, training and testing data for each task are fully isolated into independent subgraphs. Specifically, for each task \tau_{k} , the model is trained on a training set of graphs \mathcal{G}{\text{train}}^{k} and evaluated on a distinct set of unseen test graphs \mathcal{G}{\text{test}}^{k} , where \mathcal{G}{\text{train}}^{k}\cap\mathcal{G}{\text{test}}^{k}=\emptyset . This setup challenges the model to generalize to new classes as well as to entirely novel graph structures and node instances that it has never encountered during training. 3.2. Graph Neural Networks A GNN computes the hidden representation for a node h_{i}^{(l)} at a given layer l by aggregating features from its neighbors \mathcal{N}(i) ( 8 ) . This process is generally defined by applying a transformation matrix \mathbf{W}^{(l)} and a non-linear activation function \sigma(\cdot) : (1) h_{i}^{(l)}=\sigma\left(\sum_{j\in\mathcal{N}(i)}\mathcal{A}_{ij}h_{j}^{(l-1)}\mathbf{W}^{(l)}\right) Here, \mathcal{A} is a matrix defining the aggregation strategy, and h_{i}^{(0)} is the initial node feature. SGC ( 30 ) streamlines the GNN architecture by removing intermediate non-linear activation functions. This decouples the model into two distinct stages: (1) Neighborhood Aggregation The initial node features \mathbf{X} are aggregated over a K-hop neighborhood by being multiplied by the K -th power of the normalized adjacency matrix \mathbf{S} . This creates a fixed, pre-processed feature matrix \mathbf{X}^{\prime} : (2) \mathbf{X}^{\prime}=\mathbf{S}^{K}\mathbf{X} (2) Feature Transformation: A trainable Multi-Layer Perceptron (MLP) is then applied to the propagated features \mathbf{X}^{\prime} .: (3) \hat{\mathbf{Y}}=\text{softmax}\left(\sigma\left(\mathbf{X}^{\prime}\mathbf{W}_{1}\right)\mathbf{W}_{2}\right) where \mathbf{W}_{1} and \mathbf{W}_{2} are the trainable weight matrices of the MLP, and \sigma(\cdot) is a non-linear activation function, such as ReLU. 4. Methodology The proposed Task-Aware Adaptive Modulation (TAAM) framework offers a replay-free and resource-efficient solution for continual learning on graphs. TAAM utilizes a frozen Simple Graph Convolution (SGC) backbone as a stable feature extractor. For every new task in the sequence \mathcal{T}={\tau_{1},\tau_{2},\dots,\tau_{N}} , it introduces a lightweight, task-specific Neural Synapse Modulator (NSM) to dynamically adapt the backbone’s output. Each task is assigned its own NSM ( {\text{NSM}_{1}},{\text{NSM}_{2}},\dots,{\text{NSM}_{N}} ), as illustrated in Figure 2 . At inference time, the system selects the appropriate NSM using robust task-ID inference, which is based on prototypes created by the Anchored Multi-hop Propagation (AMP) technique. 4.1. Anchored Multi-hop Propagation for Task-ID Inference A critical component for stability in a continual learning setting without task labels at inference is the ability to accurately identify the task to which incoming data belongs. To achieve this, we propose generating a single, highly representative prototype vector \mathbf{p}_{k} for each task \tau_{k} . We introduce A nchored M ulti-hop P ropagation (AMP), a method designed to create robust prototypes by capturing multi-scale neighborhood information. AMP leverages the principles of Approximate Personalized PageRank (APPNP) ( 9 ) , which diffuses node features across the graph while retaining a connection to the initial features via a teleport probability \alpha . The propagation rule for the node feature matrix \mathbf{H} is: (4) \mathbf{Z}^{(i+1)}=(1-\alpha)\mathbf{S}\mathbf{Z}^{(i)}+\alpha\mathbf{H}^{(0)} where \mathbf{Z}^{(0)}=\mathbf{H}^{(0)} are the initial node features, and \mathbf{S} is the symmetrically normalized adjacency matrix. Inspired by SIGN ( 23 ) , we create a multi-scale representation by concatenating the diffused features from different propagation depths (hops). The final enriched feature matrix for prototype generation, \mathbf{H}_{\text{AMP}} , is formed as: (5) \mathbf{H}_{\text{AMP}}=\text{CONCAT}(\mathbf{Z}^{(h_{1})},\mathbf{Z}^{(h_{2})},\dots,\mathbf{Z}^{(h_{m})}) where \{h_{1},\dots,h_{m}\} is the set of selected hops. The prototype \mathbf{p}_{k} for task \tau_{k} is then the mean of these enriched features over its training nodes \mathcal{V}_{k}^{\text{train}} : (6) \mathbf{p}_{k}=\frac{1}{|\mathcal{V}_{k}^{\text{train}}|}\sum_{i\in\mathcal{V}_{k}^{\text{train}}}(\mathbf{H}_{\text{AMP}})_{i} During inference, we generate a prototype for the test data and identify the task ID by finding the most similar stored prototype using cosine similarity. The pseudocode for AMP is detailed in Algorithm 1 . Input : Graph g , Initial features \mathbf{H}^{(0)} , Training node set \mathcal{V}_{k}^{\text{train}} , AMP parameters ( L_{hops},\alpha,K ); Output : Task prototype vector \mathbf{p}_{k} ; begin \mathcal{H}_{collect}\leftarrow\emptyset ; \mathbf{Z}\leftarrow\mathbf{H}^{(0)} ; if 0\in L_{hops} then \mathcal{H}_{collect}\leftarrow\mathcal{H}_{collect}\cup\{\mathbf{Z}\} ; end if for i\leftarrow 1 to K do \mathbf{Z}\leftarrow(1-\alpha)\mathbf{S}\mathbf{Z}+\alpha\mathbf{H}^{(0)} ; if i\in L_{hops} then \mathcal{H}_{collect}\leftarrow\mathcal{H}_{collect}\cup\{\mathbf{Z}\} ; end if end for \mathbf{H}_{\text{AMP}}\leftarrow\text{CONCAT}(\mathcal{H}_{collect}) ; \mathbf{p}_{k}\leftarrow\frac{1}{|\mathcal{V}_{k}^{\text{train}}|}\sum_{v\in\mathcal{V}_{k}^{\text{train}}}(\mathbf{H}_{\text{AMP}})_{v} ; return \mathbf{p}_{k} ; end Algorithm 1 Anchored Multi-hop Propagation 4.2. Neural Synapse Modulator For each task \tau_{k} , we introduce a new, lightweight N eural S ynapse M odulator (NSM). The NSM acts as a task-specific expert that adapts the behavior of the frozen SGC backbone. It is a dynamic, node-attentive module that generates Feature-wise Linear Modulation (FiLM) ( 2 ; 22 ) parameters, \boldsymbol{\gamma} and \boldsymbol{\beta} , for each node. Specifically, the NSM for task \tau_{k} contains a unique, learnable task embedding \mathbf{e}_{k}\in\mathbb{R}^{D_{task}} . To generate modulation parameters efficiently, we use a low-rank factorization. First, a set of H base modulation vectors for H attention heads, \mathbf{M}_{k}\in\mathbb{R}^{H\times 2F^{\prime}} , is generated from the task embedding: (7) \mathbf{M}_{k}=(\mathbf{W}_{A}\mathbf{e}_{k})\cdot\mathbf{W}_{B} where \mathbf{W}_{A} and \mathbf{W}_{B} are learnable low-rank projection matrices. Next, for each node i with feature vector \mathbf{h}_{i} , we compute attention weights over the H heads: (8) \mathbf{a}_{i}=\text{softmax}(\mathbf{h}_{i}\mathbf{W}_{\text{att}}) where \mathbf{W}_{\text{att}}\in\mathbb{R}^{F\times H} . The final per-node FiLM parameters are a weighted combination of the base modulations: (9) [\boldsymbol{\gamma}_{i},\boldsymbol{\beta}_{i}]=\mathbf{a}_{i}^{\top}\mathbf{M}_{k} These parameters are then used to modulate the node’s feature vector \mathbf{h}_{i} after the SGC’s neighborhood aggregation step, followed by a residual connection: (10) \mathbf{h}^{\prime}_{i}=(\boldsymbol{\gamma}_{i}\odot\text{LayerNorm}(\mathbf{h}_{i})+\boldsymbol{\beta}_{i})+\mathbf{h}_{i} This node-attentive mechanism allows the NSM to provide fine-grained, context-specific adaptation of the backbone’s representations. Figure 3 . Performance matrices of differernt method on Products dataset.Darker colors indicate higher accuracy. TAAM (Ours) demonstrates consistently high accuracy across all tasks, indicating zero catastrophic forgetting. A heat map comparison showing performance matrices. TAAM maintains high accuracy (darker colors) along the diagonal and the lower triangle, demonstrating its ability to prevent catastrophic forgetting compared to baselines. 4.3. Unified Classifier and Training Objective To accommodate the expanding set of classes, we employ a unified linear classifier that grows dynamically. Upon the arrival of a new task \tau_{k} with class set \mathcal{C}_{k} , the classifier’s weight matrix \mathbf{W}_{k-1}\in\mathbb{R}^{D_{hid}\times C_{k-1}} is expanded to a new matrix \mathbf{W}_{k}\in\mathbb{R}^{D_{hid}\times C_{k}} , where C_{k}=C_{k-1}+|\mathcal{C}_{k}| . To preserve prior knowledge, the weights for existing classes are copied directly from \mathbf{W}_{k-1} , while only the weights corresponding to the new classes are randomly initialized. During the training phase for task \tau_{k} , we freeze all parameters except for those of the new NSM and the newly added weights in the classifier. This design strictly isolates parameter updates and prevents catastrophic forgetting at the decision layer. To mitigate class imbalance, our training objective for task \tau_{k} is a weighted cross-entropy loss. The weight for each class c\in\mathcal{C}_{k} is its inverse frequency in the training data, w_{c}=1/N_{c} , where N_{c} is the number of training instances for class c . Crucially, the loss is computed exclusively over the set of new classes \mathcal{C}_{k} : (11) \mathcal{L}_{k}=-\sum_{v\in\mathcal{V}_{k}^{\text{train}}}w_{y_{v}}\log\left(\frac{\exp(\mathbf{z}_{v,y_{v}})}{\sum_{c^{\prime}\in\mathcal{C}_{k}}\exp(\mathbf{z}_{v,c^{\prime}})}\right) where \mathcal{V}_{k}^{\text{train}} are the training nodes for task \tau_{k} , \mathbf{z}_{v} is the full logit vector for node v , and y_{v}\in\mathcal{C}_{k} is its corresponding label. Table 1 . Comparison of different models on different data sets in the GCIL setting. The bold results are the best performance excluding Joint and Oracle. A ‘ \uparrow ‘ indicates that a greater value represents better performance, while a ‘ \downarrow ‘ indicates the opposite. ‘Joint‘ can access the data of all tasks; ‘Oracle‘ can access the data and task IDs of all tasks. Category Model CoraFull Arxiv Reddit Products Photo WikiCS Computer Citeseer AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) Lower bound Finetuning 1.8 \pm 0.2 13.7 \pm 2.5 4.9 \pm 0.0 11.8 \pm 4.1 4.9 \pm 0.0 4.9 \pm 0.0 4.7 \pm 0.8 31.1 \pm 1.6 23.2 \pm 0.0 38.6 \pm 6.5 18.8 \pm 0.1 41.5 \pm 7.9 19.0 \pm 0.1 28.1 \pm 12.0 31.4 \pm 0.1 52.3 \pm 5.3 LwF(2017) 1.8 \pm 0.2 18.2 \pm 2.1 4.9 \pm 0.0 30.8 \pm 4.8 5.6 \pm 0.9 58.4 \pm 5.3 21.4 \pm 0.9 33.4 \pm 0.7 12.5 \pm 8.0 46.2 \pm 3.1 14.7 \pm 1.2 78.4 \pm 6.0 8.5 \pm 7.8 41.8 \pm 32.0 31.4 \pm 0.1 70.7 \pm 0.8 Regularisation MAS(2018) 2.8 \pm 0.3 67.4 \pm 1.4 4.8 \pm 0.0 80.9 \pm 4.2 5.0 \pm 0.1 97.5 \pm 0.1 3.9 \pm 0.3 88.7 \pm 0.4 18.6 \pm 0.1 92.3 \pm 2.0 18.6 \pm 0.1 92.3 \pm 1.8 23.1 \pm 0.2 87.3 \pm 14.2 29.9 \pm 0.1 80.9 \pm 0.4 TWP(2021) 7.2 \pm 0.9 65.7 \pm 1.3 4.9 \pm 0.0 60.5 \pm 5.3 5.0 \pm 0.3 92.8 \pm 1.5 5.5 \pm 0.2 78.3 \pm 1.7 23.2 \pm 0.1 73.8 \pm 7.5 18.6 \pm 0.1 89.2 \pm 2.5 19.2 \pm 0.3 64.7 \pm 15.4 36.5 \pm 0.4 66.9 \pm 1.2 ER-GNN(2021) 2.1 \pm 0.0 14.1 \pm 1.8 16.3 \pm 2.7 40.4 \pm 2.4 85.0 \pm 0.8 12.3 \pm 0.8 22.5 \pm 0.9 33.1 \pm 1.2 94.3 \pm 0.2 2.4 \pm 0.2 84.1 \pm 0.3 7.3 \pm 0.5 87.6 \pm 0.6 10.6 \pm 0.9 61.3 \pm 0.4 34.5 \pm 0.6 Replay SSM(2022) 60.3 \pm 0.6 11.2 \pm 0.3 7.7 \pm 2.1 8.8 \pm 2.5 91.9 \pm 0.3 1.5 \pm 0.2 32.9 \pm 1.5 15.3 \pm 1.6 92.6 \pm 0.7 1.9 \pm 0.5 74.3 \pm 1.6 4.4 \pm 1.1 88.7 \pm 0.9 3.1 \pm 0.8 47.1 \pm 4.5 6.5 \pm 2.2 PDGNNs(2024) 49.8 \pm 27.0 11.5 \pm 7.9 44.5 \pm 0.2 14.2 \pm 0.1 94.9 \pm 0.1 1.4 \pm 0.0 45.2 \pm 0.1 7.1 \pm 0.4 94.6 \pm 0.3 2.2 \pm 0.2 84.2 \pm 0.5 3.8 \pm 0.4 91.4 \pm 0.8 5.5 \pm 0.9 71.0 \pm 0.1 10.9 \pm 0.1 DeLoMe(2024) 30.5 \pm 0.1 47.0 \pm 0.5 31.0 \pm 0.2 24.1 \pm 0.2 94.1 \pm 0.6 0.2 \pm 0.1 37.8 \pm 0.5 35.8 \pm 0.2 47.4 \pm 0.2 64.9 \pm 0.3 35.6 \pm 0.3 67.1 \pm 0.3 37.1 \pm 0.1 73.7 \pm 0.2 48.4 \pm 0.2 32.3 \pm 0.2 Parameter isolation TPP(2024) 88.9 \pm 0.8 0.0 \pm 0.0 84.4 \pm 1.0 0.0 \pm 0.0 98.5 \pm 0.2 0.0 \pm 0.0 84.9 \pm 0.2 0.0 \pm 0.0 23.3 \pm 0.7 0.0 \pm 0.0 36.6 \pm 0.5 0.0 \pm 0.0 19.9 \pm 0.1 46.2 \pm 0.5 77.7 \pm 1.6 0.0 \pm 0.0 TAAM (Ours) 92.5 \pm 0.0 0.0 \pm 0.0 88.6 \pm 0.4 0.0 \pm 0.0 99.0 \pm 0.0 0.0 \pm 0.0 92.9 \pm 0.5 0.0 \pm 0.0 96.2 \pm 0.4 0.0 \pm 0.0 92.8 \pm 0.3 0.0 \pm 0.0 97.5 \pm 0.6 0.0 \pm 0.0 83.5 \pm 1.0 0.0 \pm 0.0 Upper bound Joint 64.1 \pm 0.3 - 45.2 \pm 0.1 - 95.6 \pm 0.1 - 65.2 \pm 0.3 - 95.1 \pm 0.5 - 83.7 \pm 0.3 - 94.9 \pm 0.1 - 70.7 \pm 0.1 - Oracle 94.3 \pm 0.3 - 88.5 \pm 0.3 - 99.2 \pm 0.0 - 94.0 \pm 0.6 - 97.0 \pm 0.1 - 92.9 \pm 0.2 - 97.7 \pm 0.1 - 85.2 \pm 0.2 - 5. Experiments In this section, we conduct a comprehensive empirical evaluation to validate the effectiveness and efficiency of our proposed Task-Aware Adaptive Modulation (TAAM) framework. Our experiments are designed to answer four key research questions: Q1: How does TAAM perform against state-of-the-art (SOTA) Graph Continual Learning (GCL) methods in terms of accuracy and catastrophic forgetting? Q2: How efficient is TAAM concerning computational time, memory usage, and parameter overhead? Q3: How crucial is TAAM’s task-ID inference mechanism, and how does it compare to existing approaches? Q4: What is the specific contribution of each core component within the TAAM architecture? 5.1. Experimental Setup Datasets and Baselines. We perform evaluations on eight benchmark datasets, which span several domains. Specifically, these include three citation networks: CoraFull ( 19 ) , Arxiv ( 7 ) , and Citeseer ( 31 ) ; one social network: Reddit ( 5 ) ; three product co-purchase networks: Products ( 7 ) , Photo ( 18 ) , and Computer ( 18 ) ; and one webpage-based network: WikiCS ( 12 ) , with statistics detailed in Table 2 . Our baselines cover the spectrum of GCL strategies,including regularization, replay-base, and parameter isolation leading methods. We use Joint training and Oracle models as upper bounds. Implementation Details. To ensure a rigorous and practical evaluation, our experimental setup adapts the CGL benchmark ( 34 ) to a strict inductive learning scenario . In this setting, the training, validation, and test sets are constructed from entirely disjoint subgraphs. This deliberate separation is a crucial correction to prior evaluation protocols, as it prevents any potential data leakage from the test distribution during the training phase. Consequently, this setup guarantees a more realistic and unbiased assessment of a model’s generalization capabilities in a continual learning environment. For a fair comparison across all methods, we employ a unified backbone feature extractor: a two-layer Simplifying Graph Convolutional Network (SGC) ( 30 ) with a 256-dimensional hidden layer. The SGC is configured without batch normalization, dropout, or bias. Its parameters are randomly initialized and subsequently frozen throughout the entire continual learning process, ensuring that all performance gains come from the continual learning modules themselves. All models were trained for 200 epochs per task using the Adam optimizer, with a learning rate of 0.005 and a weight decay of 5\times 10^{-4} . We use the same set of random seeds for all experiments to ensure reproducibility. Hyperparameters for all baseline methods were configured according to their original publications. For our proposed model, TAAM, the dimension of the learnable task embedding, \mathbf{e}_{\tau} , is set to 6. The internal node-attentive mechanism in each Node-Specific Modulator (NSM) uses K=2 heads. Critically, for each new task, only the parameters of the newly instantiated NSM and its corresponding classifier are trainable, embodying our lightweight learning paradigm. Evaluation Metrics. We evaluate performance using Average Accuracy (AA) and Average Forgetting (AF), where a higher AA indicates better overall accuracy and a lower AF indicates less catastrophic forgetting. Table 2 . Statistics of the datasets. Dataset Nodes Edges Feats. Classes Tasks Citeseer 3,327 9,228 3,703 6 3 Photo 48,362 500,928 745 8 4 WikiCS 11,701 431,726 300 10 5 Computer 13,752 491,722 767 10 5 CoraFull 19,793 130,622 8,710 70 35 Arxiv 169,343 1,166,243 128 40 20 Reddit 232,965 114,615,892 602 40 20 Products 2,449,029 61,859,036 100 47 23 5.2. Main results State-of-the-Art Performance ( Q1 ) As shown in Table 1 , TAAM establishes a new state-of-the-art by achieving the highest Average Accuracy (AA) across all eight datasets while simultaneously eliminating catastrophic forgetting (0.0% AF). This superiority stems from its unique architectural design, which overcomes the limitations of prior paradigms. Unlike replay-based methods, TAAM avoids their inherent computational and privacy overhead while still delivering dominant performance on large-scale graphs like Products (92.9% vs. 45.2% for PDGNNs). More critically, its advancement over other parameter-isolation methods like TPP is driven by its novel Task-Aware module and fine-grained Neural Synapse Modulators (NSMs). This mechanism provides superior adaptability, leading to remarkable performance gains such as the 77.6% improvement on the Computer dataset. Ultimately, TAAM presents a more effective solution to the stability-plasticity trade-off : its core framework guarantees perfect stability (zero forgetting), while its adaptive modulators deliver SOTA plasticity. Visualizing Knowledge Retention To provide a more intuitive understanding of this zero-forgetting property, we visualize the model’s task-wise performance after the final training phase. As depicted in Figure 3 , the heatmaps vividly illustrate our method’s superior knowledge retention. The matrix for TAAM (Ours) exhibits a consistently dark and uniform color across the lower triangle. This signifies both excellent plasticity (the dark diagonal, indicating high accuracy on newly learned tasks) and perfect stability (no color degradation in the off-diagonal elements, confirming zero forgetting). In stark contrast, the heatmaps for replay-based methods like DeLoMe and regularization-based ER-GNN reveal a clear pattern of knowledge decay. While their diagonals indicate an ability to learn, the pronounced color fading down each column provides clear visual evidence of catastrophic forgetting, corroborating the high AF% values reported in Table 1 . 5.3. Efficiency Analysis Efficiency Analysis ( Q2 ) A core design philosophy of TAAM is resource efficiency, a critical factor for practical GCL deployment. We analyze this through the lens of the performance-cost trade-off, considering parameters, memory, and time. As illustrated in Figure 5 , the optimal efficiency profile—high accuracy with low computational cost—resides in the top-left quadrant. TAAM is consistently positioned in this optimal region , achieving state-of-the-art accuracy with minimal running time and memory footprint. However, this plot only reveals part of the story. For instance, while TPP appears competitive in inference time, this view conceals the heavy computational overhead of its mandatory pre-training phase . Similarly, some methods are omitted entirely for clarity due to their extreme inefficiency; DeLoMe , for example, requires substantial time to generate node prototypes, with a single experiment on the large-scale Reddit dataset consuming approximately one hour. Parameter Efficiency This computational efficiency is rooted in our model’s design. As shown in Table 4 , TAAM demonstrates exceptional parameter efficiency by introducing only 0.02M - 3.08M additional parameters per task (measured by the number of float32 values), which is orders of magnitude lower than methods like ER-GNN and PDGNNs. This makes TAAM a truly lightweight solution. For scenarios where peak performance is paramount, our variant TAAM-L offers a flexible trade-off, further boosting accuracy with only a modest increase in parameters (Table 3 ), demonstrating the adaptability of our framework to different efficiency-performance requirements. Figure 4 . Task ID prediction accuracy of AMP and Laplacian smoothing (LS) of TPP A bar chart comparing the task ID prediction accuracy of our proposed Anchored Multi-hop Propagation (AMP) method against the Laplacian smoothing (LS) baseline. Table 3 . Ablation results of TAAM and its variants. Ablated Components CoraFull Arxiv Reddit Products Photo WikiCS Computer Citeseer AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) AA%( \uparrow ) AF%( \downarrow ) w/o Task-Aware 2.0 \pm 0.3 5.0 \pm 0.9 4.9 \pm 0.0 14.8 \pm 2.4 5.3 \pm 0.9 27.9 \pm 4.6 2.7 \pm 0.5 11.6 \pm 3.0 21.2 \pm 0.8 55.6 \pm 10.5 18.9 \pm 0.1 32.1 \pm 14.2 18.9 \pm 0.1 32.3 \pm 14.6 29.7 \pm 0.6 37.4 \pm 14.4 w/o NSMs 5.8 \pm 0.8 0.0 \pm 0.0 7.8 \pm 2.8 0.0 \pm 0.0 9.0 \pm 3.7 0.0 \pm 0.0 8.8 \pm 2.0 0.0 \pm 0.0 30.6 \pm 7.2 0.0 \pm 0.0 23.5 \pm 9.1 0.0 \pm 0.0 25.5 \pm 9.0 0.0 \pm 0.0 31.7 \pm 4.5 0.0 \pm 0.0 w/o Unified Classifier 92.2 \pm 0.5 0.0 \pm 0.0 88.3 \pm 0.4 0.0 \pm 0.0 99.0 \pm 0.0 0.0 \pm 0.0 93.0 \pm 0.5 0.0 \pm 0.0 95.7 \pm 0.7 0.0 \pm 0.0 92.9 \pm 0.2 0.0 \pm 0.0 97.7 \pm 0.5 0.0 \pm 0.0 82.0 \pm 0.8 0.0 \pm 0.0 TAAM 92.5 \pm 0.5 0.0 \pm 0.0 88.6 \pm 0.4 0.0 \pm 0.0 99.0 \pm 0.0 0.0 \pm 0.0 92.9 \pm 0.5 0.0 \pm 0.0 96.2 \pm 0.4 0.0 \pm 0.0 92.8 \pm 0.3 0.0 \pm 0.0 97.5 \pm 0.6 0.0 \pm 0.0 83.5 \pm 1.0 0.0 \pm 0.0 TAAM-L (w/o LoRA) 93.6 \pm 0.3 0.0 \pm 0.0 87.9 \pm 0.5 0.0 \pm 0.0 99.1 \pm 0.1 0.0 \pm 0.0 93.0 \pm 0.6 0.0 \pm 0.0 95.9 \pm 0.7 0.0 \pm 0.0 93.4 \pm 0.2 0.0 \pm 0.0 98.0 \pm 0.4 0.0 \pm 0.0 84.9 \pm 0.3 0.0 \pm 0.0 Table 4 . Parameter efficiency comparison. Method Total Avg AA( \uparrow ) Additional Cost Min. Max. Oracle 93.6 2.12 M 82.5638 M ER-GNN 56.7 1.80 M 103.2285 M SSM 61.9 0.30 M 51.6099 M DeLoMe 45.2 0.18 M 34.4037 M PDGNNs 71.9 1.06 M 100.2838 M TPP 64.3 0.01 M 1.7949 M TAAM-L 93.2 0.05 M 9.1812 M TAAM 92.9 0.02 M 3.0861 M 5.4. Analysis of Core Components Superior Task-ID Inference ( Q3 ) Reliable task-ID inference is the foundation of our replay-free approach. We compare our Anchored Multi-hop Propagation (AMP) against the Laplacian smoothing (LS) method used in TPP. Figure 4 shows that AMP achieves near-perfect task-ID accuracy across diverse datasets, significantly outperforming LS. This robust inference capability prevents error propagation and is key to TAAM’s stability and consistent performance. Ablation Study ( Q4 ) To dissect the model architecture, we conducted a comprehensive ablation study, The key findings from our ablation study are as follows: • w/o Task-Aware Module : The removal of this component leads to a catastrophic collapse in performance. This confirms that the task-aware module is the cornerstone of our architecture, enabling the frozen GNN backbone to effectively adapt to diverse task distributions. • w/o NSMs : Excluding the NSMs results in a severe degradation in accuracy, even though forgetting is not induced. This validates our hypothesis that fine-grained, instance-level adaptation is crucial for capturing the unique characteristics of nodes within each task. • w/o Unified Classifiers : Compared with using independent classifiers for each task, adopting a unified classifier will lead to a persistent slight decline in performance, which indicates that decoupled classification heads can effectively avoid interference between tasks.. • TAAM vs. TAAM-L (w/o LoRA) : A core design choice in our framework is the integration of LoRA to parameterize the NSMs. This is a deliberate trade-off: its primary purpose is to ensure that the number of new parameters per task remains minimal, preventing the model size from growing uncontrollably as more tasks are learned. To validate this strategy, we compare our full TAAM model against TAAM-L, a variant employing a simpler modulator without the LoRA parameterization. The results show that complete model is slightly better than that of the variant without LoRA It has been proven that LoRA has achieved an excellent balance between ensuring parameter efficiency and maintaining model performance. 5.5. Parameter Efficiency Finally, we analyze the parameter efficiency of TAAM. Table 4 compares the additional parameter cost required by different continual learning methods. It is evident that TAAM and its variant TAAM-L are exceptionally parameter-efficient. Our method requires only 0.02M to 3.08M additional parameters, a cost that is orders of magnitude lower than most replay-based and regularization-based methods like ER-GNN and PDGNNs, and even more efficient than the parameter-isolation baseline TPP. This demonstrates that TAAM not only achieves superior performance but does so with remarkable efficiency, making it a practical and scalable solution for real-world graph continual learning scenarios. Figure 5 . Comparison of running time and AA A scatter plot or line chart comparing the training time efficiency and Average Accuracy (AA) between TAAM and other state-of-the-art graph continual learning methods. 6. Conclusion In this paper, we introduce TAAM, a new replay-free framework for Graph Continual Learning (GCL) designed to address the stability-plasticity dilemma without the memory and privacy issues associated with replay-based methods. The core of TAAM balances performance and parameter efficiency: it combines a frozen GNN backbone, which maintains prior knowledge, with lightweight, task-specific Neural Synapse Modulators (NSMs) that are trained and kept fixed for each new task. This design intentionally adds only a small number of parameters per task, preventing the model size from growing uncontrollably while still allowing enough plasticity to learn new concepts. Additionally, we propose Anchored Multi-hop Propagation (AMP), a theoretically grounded method to infer task IDs in real-world scenarios where they are unknown. Importantly, we identify flaws in existing GCL benchmarks and perform a thorough evaluation in a more rigorous and realistic inductive learning setting. Extensive experiments across eight datasets show that TAAM consistently and significantly outperforms state-of-the-art methods. Appendix A Theoretical Analysis of AMP In this section, we provide a theoretical analysis to ground our choice of Anchored Multi-hop Propagation (AMP) for task prototype generation. We aim to demonstrate that AMP is inherently designed to produce task prototypes with high inter-task separability and low intra-task variance. This property is crucial for minimizing the task classification error during inference, which is the cornerstone of our model’s stability. A.1. Objective Formulation Let us consider two distinct tasks, \tau_{k} and \tau_{j} , with their respective training node sets \mathcal{V}_{k}^{\text{train}} and \mathcal{V}_{j}^{\text{train}} , and initial feature matrices \mathbf{H}_{k}^{(0)} and \mathbf{H}_{j}^{(0)} . Our goal is to generate prototypes \mathbf{p}_{k} and \mathbf{p}_{j} that are maximally separable. An ideal prototype generation function, f(\cdot) , should achieve two objectives: (1) Minimize Intra-Task Variance: The features of nodes within the same task should be compactly clustered around their task’s prototype. We can define the intra-task variance for task \tau_{k} as: (12) V_{\text{intra}}(\tau_{k})=\mathbb{E}_{v\in\mathcal{V}_{k}^{\text{train}}}\left[||f(\mathbf{H}^{(0)})_{v}-\mathbf{p}_{k}||^{2}\right] where f(\mathbf{H}^{(0)})_{v} is the transformed feature of node v . (2) Maximize Inter-Task Distance: The prototypes of different tasks should be far apart from each other in the embedding space. We define the squared inter-task distance as: (13) D_{\text{inter}}(\tau_{k},\tau_{j})^{2}=||\mathbf{p}_{k}-\mathbf{p}_{j}||^{2} We will now analyze how each component of AMP contributes to optimizing these two objectives. A.2. Intra-Task Variance Reduction via Graph Diffusion The core of AMP is the APPNP propagation rule (Eq. (8)). Let’s analyze the iterative diffusion component, \mathbf{Z}^{(i+1)}\propto\mathbf{S}\mathbf{Z}^{(i)} . This operation is a form of graph Laplacian smoothing. In spectral graph theory, multiplying features by the normalized adjacency matrix \mathbf{S} is equivalent to applying a low-pass filter to the graph signal (i.e., the node features). A low-pass filter attenuates high-frequency components of the signal. In the context of graphs, high-frequency components correspond to large feature differences between adjacent nodes. Therefore, the diffusion process forces the features of connected nodes to become more similar. This process is equivalent to minimizing the graph’s Dirichlet energy: (14) \mathcal{E}(\mathbf{Z})=\frac{1}{2}\sum_{u,v}A_{uv}\left|\left|\frac{\mathbf{z}_{u}}{\sqrt{d_{u}}}-\frac{\mathbf{z}_{v}}{\sqrt{d_{v}}}\right|\right|^{2}=\text{Tr}(\mathbf{Z}^{\top}(\mathbf{I}-\mathbf{S})\mathbf{Z}) where \text{Tr}(\cdot) is the trace of a matrix. Iteratively applying \mathbf{S} minimizes this energy, effectively smoothing the features across the graph. Our key assumption, standard in graph-based learning, is that intra-task connectivity is significantly denser than inter-task connectivity . Under this assumption, the feature smoothing will primarily occur *within* the node set of a given task \tau_{k} . Consequently, the features of nodes in \mathcal{V}_{k}^{\text{train}} will converge towards a common representation, leading to a direct reduction in the intra-task variance, V_{\text{intra}}(\tau_{k}) . The resulting prototype \mathbf{p}_{k} , being the mean of these smoothed features, becomes a more stable and robust representation of the entire task. A.3. Preserving Inter-Task Separability While feature smoothing is beneficial for reducing intra-task variance, excessive smoothing (oversmoothing) is detrimental, as it would cause all nodes in the graph to converge to the same vector, collapsing all prototypes and making D_{\text{inter}}\to 0 . AMP employs two mechanisms to counteract this and preserve separability. 1. Anchoring with Teleport Probability: The teleport term \alpha\mathbf{H}^{(0)} in Eq. (8) acts as a crucial regularizer. At each propagation step, it pulls the smoothed features back towards their initial, task-specific representations. This "anchoring" ensures that the feature distributions of different tasks, which are assumed to be initially distinct, do not completely merge. It preserves the global, task-level differences while the diffusion term smooths out local, intra-task variations. 2. Multi-scale Feature Concatenation: Different tasks may be characterized by different topological scales. For instance, one task’s graph may consist of many small, tight clusters, while another’s may be defined by long-range dependencies. A prototype based on a single propagation depth K might fail to capture these diverse structural fingerprints. By concatenating features from multiple hops \{h_{1},\dots,h_{m}\} (Eq. (9)), \mathbf{H}_{\text{AMP}} creates a far richer and more discriminative embedding. This allows the prototype to simultaneously encode local patterns (from small h_{i} ) and global community structure (from large h_{i} ), significantly increasing the likelihood that the prototypes for structurally or featurally distinct tasks will be distant from each other, thus maximizing D_{\text{inter}} . A.4. Minimizing Task Classification Error The task-ID inference is a nearest-centroid classification problem based on cosine similarity, which is closely related to Euclidean distance for normalized vectors. The probability of misclassifying a test prototype \mathbf{p}_{\text{test}} belonging to task \tau_{k} as belonging to task \tau_{j} is a function of the distance to their respective centroids and the variance of the features. For a simplified two-class case, the Bayes classification error rate P_{e} is bounded by a function that decreases as the ratio of inter-task distance to intra-task variance increases: (15) P_{e}(\tau_{k}\to\tau_{j})\leq\exp\left(-\frac{D_{\text{inter}}(\tau_{k},\tau_{j})^{2}}{8\cdot\max(V_{\text{intra}}(\tau_{k}),V_{\text{intra}}(\tau_{j}))}\right) From our analysis, AMP is explicitly designed to optimize this ratio. The graph diffusion component minimizes the denominator ( V_{\text{intra}} ), while the anchoring and multi-scale concatenation mechanisms work to preserve or maximize the numerator ( D_{\text{inter}} ). By generating prototypes with these desirable geometric properties, AMP provides a strong foundation for a classifier that can robustly distinguish between tasks, thereby minimizing the task-ID prediction error and ensuring the overall stability of the continual learning system. Datasets and Baselines. We perform evaluations on eight benchmark datasets, which span several domains. Specifically, these include three citation networks: CoraFull ( 19 ) , Arxiv ( 7 ) , and Citeseer ( 31 ) ; one social network: Reddit ( 5 ) ; three product co-purchase networks: Products ( 7 ) , Photo ( 18 ) , and Computer ( 18 ) ; and one webpage-based network: WikiCS ( 12 ) , with statistics detailed in Table 2 . Our baselines cover the spectrum of GCL strategies,including regularization, replay-base, and parameter isolation leading methods. We use Joint training and Oracle models as upper bounds. Appendix B Details of Datesets • Citeseer ( 31 ) : Another widely-used citation network where nodes are scientific publications and edges are citations. The papers are classified into six categories. Similar to Cora, node features are based on the document’s textual content. • CoraFull ( 19 ) : A larger citation network encompassing 70 classes. Nodes correspond to academic papers and edges denote citation links between them. • Arxiv ( 7 ) : This dataset is constructed from the collaboration network of Computer Science (CS) papers on the arXiv platform, indexed by MAG ( 24 ) . Nodes, which represent CS papers, are classified into 40 subject areas. Node features are generated by averaging the word embeddings of the titles and abstracts. • Reddit ( 5 ) : A social network dataset derived from Reddit posts made in September 2014. Here, nodes are individual posts, and an edge connects two posts if the same user has commented on both. The objective is to classify posts into their corresponding communities (subreddits). Node features are engineered from post attributes such as title, content, score, and comment count. • Products ( 7 ) : An Amazon product co-purchase network where nodes represent products and edges indicate that two products are frequently bought together. Node features are derived from dimensionality-reduced bag-of-words representations of the product descriptions. • WikiCS ( 12 ) : A dataset constructed based on computer science articles from Wikipedia. Nodes represent articles, and edges are hyperlinks between them. The task is to classify the articles into 10 categories based on their textual content. • Computer ( 18 ) : A dataset from Amazon’s group-buying network. Nodes represent goods related to computers, while edges indicate the relationship where goods are purchased together. The task is to classify them into their respective categories. The node features are derived from product reviews. • Photo ( 18 ) : Similar to the Computer dataset, this is also part of Amazon’s group-buying network. Nodes are goods related to photography (such as cameras and lenses), while edges connect goods that are usually purchased together. Node features are generated by the bag-of-words representation of product reviews. Appendix C Details of baselines We evaluate our method on a suite of leading CGL methods. including three mainstream categories based on Replay, regularization, and parameter isolation. • Finetuning ( 8 ) is denotes the backbone GNN without continual learning technique. Therefore,this can be viewed as the lower bound on the continual learning performance. • LwF ( 10 ) employs knowledge distillation to maintain performance on old tasks. It minimizes the discrepancy between the output logits of the previous and current models, effectively transferring knowledge without retaining old data. • MAS ( 1 ) also preserves vital parameters from past tasks, but bases the importance on the sensitivity of the model’s predictions to parameter shifts. • TWP ( 13 ) focuses on preserving important parameters within a topological aggregation framework and introduces a loss minimization strategy for previous tasks that relies on these preserved parameters. • ER-GNN ( 38 ) is a replay-based method that constructs a memory buffer by storing representative nodes selected from previous tasks. • SSM ( 35 ) enhances graph continual learning by explicitly incorporating topological information. It stores sparsified computation subgraphs of selected nodes in its memory. • PDGNNs ( 33 ) decouple trainable parameters from computational subgraphs through topology-aware embeddings (TEs). These TEs are compressed representations of ego-subnetworks that significantly reduce memory overhead. • DeLoMe ( 21 ) constructed a memory composed of lossless node prototypes, aiming to fully capture and store the graph information of old tasks. Meanwhile, to address the category imbalance that occurs when new and old data are mixed, this method further proposes a de-biased GCL loss function. • TPP ( 20 ) a GCL method that avoids both data replay and catastrophic forgetting, leverages a Laplacian smoothing technique for precise task-ID inference and employs graph prompts to dynamically adapt the GNN into specialized modules for each task. • Joint is base the backbone SGC can can get access to the data of all tasks, representing a theoretical upper bound of graph class-incremental learning • Oracle is base the backbone SGC can can get access to the data of all tasksand task IDs, representing a stronger upper bound of graph class-incremental learning. Figure 6 . Performance comparison of our method, TAAM (Ours, purple line), against leading CGL methods on four datasetsin inductive learning scenario. Figure 7 . Performance comparison of our method, TAAM (Ours, purple line), against leading CGL methods on other four datasets in inductive learning scenario. Appendix D Implementation Details All experiments are conducted on a server equipped with an Intel(R) Xeon(R) Platinum 8368Q CPU @ 2.60GHz, 256 GB of RAM, and a single NVIDIA L40 GPU. The software stack includes Python 3.8.16, PyTorch 1.10.1, DGL 0.9.0, OGB 1.3.6, and CUDA 12.4. Appendix E Evaluation Metrics To comprehensively evaluate the model’s performance in a continual learning setting, we adopt two standard metrics: Average Accuracy and Average Forgetting . Our evaluation protocol is based on the accuracy matrix M\in\mathbb{R}^{T\times T} , where T represents the total number of tasks. An element M_{t,j} of this matrix denotes the classification accuracy on task j after the model has been sequentially trained up to task t . Average Accuracy (AA) measures the overall performance of the model across all tasks after the entire training sequence is complete. It is calculated by averaging the accuracies on all tasks j after the final task T has been learned. A higher AA signifies a better-performing model. The formula is: \text{AA}=\frac{\sum_{j=1}^{T}M_{T,j}}{T} Average Forgetting (AF) quantifies how much the model forgets previously learned tasks as it acquires new knowledge. It is computed by averaging the difference between the peak accuracy on a task (i.e., immediately after it was learned) and its accuracy after training on a subsequent task, averaged over all past tasks. The formula is given by: \text{AF}=\frac{\sum_{j=1}^{T-1}(M_{T,j}-M_{j,j})}{T-1}