Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Continual Learning and Catastrophic Forgetting

Abstract. Continual learning studies the acquisition and revision of competence as experience arrives sequentially and historical training data become restricted or unavailable. Catastrophic forgetting is a marked reduction in performance on previously learned tasks after subsequent training. This fundamental research monograph examines historical connectionist evidence, complementary-learning-systems accounts, representational and gradient interference, stability–plasticity, task-, domain-, and class-incremental experimental designs, and methodological definitions of retention, transfer, and intransigence. It critically synthesizes primary research on experience replay, generative rehearsal, synaptic consolidation, parameter importance, architecture expansion, exemplar management, knowledge distillation, and prompt-based adaptation. Comparative attention is given to benchmark validity, competing failure explanations, online resource limits, reinforcement-learning agents, and recent studies of continual adaptation in language models. The resulting synthesis establishes which scientific claims can be supported by controlled sequential-learning measurements and identifies open problems concerning longevity, external memory, safety, selective forgetting, and functional continuity in configured artificial systems. This is a scholarly narrative review, not an original experimental study, registered systematic review, or meta-analysis.

Keywords: continual learning; catastrophic forgetting; catastrophic interference; sequential learning; stability–plasticity dilemma; elastic weight consolidation; experience replay; generative replay; incremental learning; synaptic intelligence; class-incremental learning; gradient interference; artificial memory; lifelong learning

The Scientific Problem of Sequential Learning

Continual learning describes the acquisition, revision, and deployment of competence when experience arrives sequentially and earlier training examples may become inaccessible. The empirical object is a learner evaluated at multiple points in its history, under distributions that can change in classes, goals, contexts, or observations. Catastrophic forgetting is a marked deterioration of previously acquired performance after subsequent learning. A proper diagnosis requires evidence that the earlier capability was learned, that later training intervened, and that evaluation conditions remain interpretable. The field connects experimental cognitive psychology with machine learning because learning history and interference are central to both. The modern classification literature organizes the problem through explicit protocols rather than a single universal forgetting score De Lange et al. (2022) Wang et al. (2024).

Sequential adaptation raises four linked questions. What is retained when the environment changes? Which changes in a shared representation undermine an old competence? How can the system acquire genuinely new information without exhausting its available capacity? What measurements demonstrate durable improvement rather than success on a recently rehearsed test? An algorithm may preserve yesterday's accuracy by refusing to adapt, or may master a new distribution while corrupting earlier decision boundaries. Those outcomes reveal a retention–acquisition tension that must be examined jointly. The methodological demand is therefore longitudinal: record performance before and after each learning episode, evaluate both old and new conditions, and report the resources and information available to the learner Chaudhry et al. (2018) Lopez-Paz & Ranzato (2017).

This monograph is a critical narrative synthesis of original connectionist experiments, cognitive and neuroscientific models, algorithmic research, benchmark comparisons, and recent analyses of large language models. It addresses forgetting in parameters, outputs, representations, and deployed behavior, with explicit attention to what each measure establishes. It evaluates regularization, replay, architectural separation, distillation, and prompt-based adaptation as solutions with different operational assumptions. The synthesis does not claim new experimental data or a registered systematic-review protocol. Its purpose is to establish the scientific boundaries, reproduce the essential argumentative structure of foundational studies, and derive testable design and evaluation requirements for artificial systems operating through the Artificial Era.

Origins in Connectionist Cognitive Science

The problem received a precise connectionist formulation when Michael McCloskey and Neal Cohen examined what happens as networks learn successive sets of items. Their 1989 chapter, published in Psychology of Learning and Motivation, showed that updates supporting a newly presented problem can rapidly disrupt an earlier trained mapping. The relevant issue was not simply that a model had finite storage. Distributed weights participated in many learned associations, so training one set of relationships changed the parameters responsible for others. This established catastrophic interference as a testable property of learning dynamics and representation, not merely an evocative description of forgetting McCloskey & Cohen (1989).

Roger Ratcliff examined recognition-memory connectionist models against constraints imposed by sequential study and subsequent retrieval. In the 1990 Psychological Review study, previously learned items deteriorated rapidly while a second problem concerned the quality of discrimination between studied and novel material. Testing different configurations did not remove the central limitations of the evaluated models. The study matters because successful fit to one psychological outcome did not license general claims about memory: a credible model needed to reproduce learning and forgetting functions together. It also placed interference within a recognizably psychological tradition of evaluating models by their behavior over time Ratcliff (1990).

Robert French's 1999 analysis consolidated the finding that distributed internal codes combine powerful generalization with vulnerability to overwriting. Similarity across tasks can facilitate transfer, yet extensive overlap can cause damaging interactions during gradient-based updating. The historical lesson is methodological: catastrophic forgetting emerges from the conjunction of learning rule, representation, sequence, and evaluation, rather than from a fixed property shared by every artificial network. Historical explanations often invoked how natural memory preserves experience, leading to distinct proposals for sparse encoding, slow consolidation, and rehearsal. These connections remain scientifically productive when analogies between brains and algorithms are accompanied by explicit measurements of mechanisms French (1999) McClelland et al. (1995).

Learning, Memory, Interference, and Retention

In cognitive psychology, interference refers to disruption associated with competing material, whereas retention concerns the availability or expression of acquired information at later testing. Proactive interference can make earlier learning interfere with later acquisition; retroactive interference can make later material impair earlier performance. Neural-network catastrophic forgetting is most directly aligned with a severe form of retroactive impairment produced by further parameter updates. That alignment is useful for generating experiments, although behavioral similarity does not identify an equivalent biological mechanism. Both domains require separating loss of stored information, retrieval failure, changed cue interpretation, and altered response policy. The same failing score can arise through each of these pathways Ratcliff (1990) French (1999).

The persistence of a trained behavior is relational: it depends on the learner, environmental input, response space, and retrieval conditions. A classifier may retain internal features associated with an old class but fail when the output layer is replaced; an agent may recall a rule while selecting another action because its objectives changed. Performance measured only on an aggregate average can hide such distinctions. A rigorous experiment records task-wise outcomes and probes representation where possible, then asks whether recovery is possible using cues, calibrated readouts, brief retraining, or restored context. Such interventions help determine whether apparently forgotten competence persists in another functional form, without asserting more than the evidence permits Wang et al. (2024).

Natural human memory exhibits forgetting and reconstruction, but its ordinary retention dynamics should not be idealized as perfect preservation. Multiple systems, reconsolidation, interference, and motivated retrieval make biological performance context-sensitive. The complementary-learning-systems framework explicitly proposes fast acquisition of episodes and slower extraction of common structure by distributed cortical representations; that account predicts why interleaving old and new examples can improve retention McClelland et al. (1995). For machine research, the important implication is experimental: compare uninterrupted sequential updating with controlled interleaving, and evaluate retention after delays and distribution changes. The analogy supports hypotheses about scheduling and architecture; it does not demonstrate neural equivalence.

A Formal Model of the Continual-Learning Objective

A minimal formalization defines a sequence of learning stages indexed by time t. At stage t the learner receives a training distribution D_t, updates parameters theta_t from theta_(t-1), and is assessed on one or more held-out distributions D_i for earlier stages. Let A_(t,i) denote performance on evaluation task i after training stage t. The matrix of such observations distinguishes acquisition on the new task, changes in retained competence, and transfer across tasks. The same protocol must specify whether identities and task boundaries are supplied to the learner, whether examples from D_i can be stored, and what resource limits apply. Without these specifications, numerical comparisons conflate different problems Lopez-Paz & Ranzato (2017) van de Ven et al. (2022).

Average accuracy after stage t can be written as the arithmetic mean of A_(t,i) across the tasks included at that stage, when scores are comparable. One operational measure of forgetting for earlier task i is the largest accuracy previously reached on that task minus A_(t,i); averaging these differences across old tasks yields a stage-specific retention summary. Backward transfer instead compares later and reference performance on earlier tasks and may be positive when later learning helps them. Forward transfer asks whether accumulated experience facilitates competence on a newly encountered problem. Authors choose different reference points and normalization conventions, so every report should print its actual formula and evaluation population Lopez-Paz & Ranzato (2017) Chaudhry et al. (2018).

The ideal joint risk would evaluate the current model on present and previous distributions, but continual settings often remove access to the latter. A training objective must then approximate information about earlier tasks by stored exemplars, model-generated experiences, protected parameters, distillation constraints, or separated components. This turns a conceptual dilemma into an experimentally tractable design space: what information about the past is represented, how reliably is it retrieved, and how much adaptation remains possible? A finite benchmark can always reward a method that matches its assumptions, making resource accounting and task availability integral to conclusions about generality Wang et al. (2024) De Lange et al. (2022).

Task, Domain, and Class Incremental Settings

Task-incremental learning presents a series of distinguishable problems and provides task identity at evaluation, allowing the correct output head or route to be selected. Domain-incremental learning retains a basic output problem while varying the input distribution, often without a task label at inference. Class-incremental learning expands the set of possible classes and requires classification across all classes encountered so far without access to the current task identity. These distinctions change both the information available during testing and the competition among predictions. The comparative experiments and formal taxonomy of van de Ven, Tuytelaars, and Tolias demonstrate that the same strategy can succeed under one scenario and struggle under another van de Ven et al. (2022).

A popular task-incremental demonstration permutes the pixels of a digit image differently for each task. Another divides digit or object categories into successive class subsets. Such experiments provide controlled distribution shifts, but can also reveal unintended shortcuts: task labels eliminate an inference problem; disjoint class heads reduce cross-class competition; and repeated exposure to simple training data may inflate performance. A method evaluated on task-labeled permuted digits cannot thereby be presumed effective for open-ended class growth or ambiguous naturalistic streams. The strong-baseline analysis by Hsu and colleagues identifies scenario definition and benchmark parity as prerequisites for meaningful rankings Hsu et al. (2018).

Real deployment mixes abrupt and gradual drift, recurrences of earlier contexts, changing class prevalences, novel labels, and transitions that have no externally announced boundaries. Stream order and access policy must be specified because these constraints directly alter the challenge. In online continual learning, a sample may be processed once; in offline incremental learning, each new stage may permit multiple passes. A model that assumes a clearly signaled task boundary may fail when the signal vanishes. The appropriate scientific question is which forms of retention and adaptation remain reliable across declared assumptions, not whether one method owns a universal position in a leaderboard Buzzega et al. (2020) Wang et al. (2024).

Distributed Representations and Gradient Interference

Artificial networks commonly distribute features across parameter sets that participate in multiple predictions. Updating parameters to reduce the error for current examples can change predictions for earlier examples that are no longer present. The resulting interference depends on representational overlap, optimization direction, model capacity, and the relation between objectives. Two tasks can share useful features while still generating conflicting gradients in particular regions of parameter space. An experiment must therefore distinguish representational similarity from destructive update: common features are a potential foundation of transfer as well as a site of competition. The classical connectionist studies established this joint dependence of generalization and interference McCloskey & Cohen (1989) French (1999).

Gradient alignment provides a local diagnostic. When updates driven by current examples increase the old-task loss, the optimization trajectories conflict in the region being traversed. This is a property of the particular losses, weights, and samples, not a timeless statement that two knowledge domains are incompatible. Gradient Episodic Memory explicitly uses stored examples and constrained updates to limit destructive interactions, providing a computational mechanism and a way to examine whether backward transfer remains possible Lopez-Paz & Ranzato (2017). Such gradient-level reasoning becomes less decisive when the learner changes its representation, optimizer state, or task structure, since an initially favorable local update can create subsequent interference.

Interference can be studied through controlled ablations. Hold the task sequence constant while varying representational overlap, network width, learning rate, feature freezing, and access to earlier data; then measure old and new losses after each stage. A good analysis also records whether adaptation is concentrated in later classification layers or shifts features throughout the representation. Evidence that a readout has changed identifies a different failure from evidence that shared features have degraded. Strong conclusions about mechanism require this localization or an intervention with predicted effects, rather than a final average-accuracy comparison alone Mirzadeh et al. (2020) Ratcliff (1990).

The Stability–Plasticity Problem

The stability–plasticity problem concerns how a learner retains consequential structure while remaining able to change when the environment demands it. A completely frozen model preserves its parameter state but cannot learn newly required distinctions through weight updates. An unconstrained optimizer may adapt readily while undermining earlier functions. Practical solutions use selective constraints, alternative memory pathways, or additional representational capacity. The tradeoff is measurable only when both retention and new learning are recorded; protecting old weights can improve one axis while creating rigidity on the other. The surveys of De Lange and colleagues and Wang and colleagues frame continual learning as joint optimization of retention, generalization, and efficiency De Lange et al. (2022) Wang et al. (2024).

An important complementary failure is intransigence: reduced ability to acquire new classes or solve a newly introduced problem despite impressive preservation of earlier performance. Chaudhry and colleagues developed evaluation measures targeting both forgetting and intransigence in incremental recognition. Their methodological point is broad. A model that shows little measured forgetting may have been constrained so tightly that it never learned the new material, while a highly plastic model may appear successful because evaluation gives most weight to current classes. A responsible study reports the entire task-performance matrix, the final evaluation across all classes, and learning success relative to relevant reference systems Chaudhry et al. (2018).

The stability–plasticity relation is not exhausted by a single numerical balance. Structural capacity, data order, duration of exposure, compute, memory size, and uncertainty about changes in the environment also determine what can be learned. Selectively forgetting obsolete associations may be desirable when earlier rules have become wrong; retaining every historical response can be harmful under concept drift. This suggests a research distinction between unintended loss of still-valid competence and deliberate revision of outdated information. Evaluation requires a specification of which earlier behaviors should remain valid, which should be updated, and whose criterion establishes that distinction Aljundi et al. (2018) Wang et al. (2024).

Complementary Learning Systems and Biological Models

The complementary-learning-systems theory of McClelland, McNaughton, and O'Reilly develops a computational explanation of why biological learning can combine rapid acquisition of episodes with slower extraction of regularities. In the proposed organization, hippocampal processes support rapid encoding of particular experiences and their reinstatement, while distributed cortical learning accumulates statistical structure through many interleaved exposures. The theory addresses interference by separating learning regimes with different representational and temporal properties. The original 1995 analysis is an explanatory model supported by psychological and neurobiological considerations; translating it into an artificial replay algorithm is a distinct computational design decision McClelland et al. (1995).

The significance of interleaving is that earlier experiences can influence changes made while later material is acquired. A system trained only on a new dataset lacks direct reminders of previously important decision boundaries; a system that reactivates old experiences during adaptation can maintain constraints distributed across several tasks. This idea predicts both benefits and limits. Rehearsal must sample informative older cases, and a narrow replay distribution may preserve familiar examples while failing to represent forgotten tails or rare contexts. Replay also entails storage, privacy, and computational costs. Thus the biological theory motivates a family of testable hypotheses rather than providing a ready-made implementation or proof of general equivalence McClelland et al. (1995) Shin et al. (2017).

Research that invokes the brain should specify whether its claim is about functional similarity, algorithmic inspiration, or homologous mechanisms. Human and animal memory involves multiple neural systems, offline activity, and contextual retrieval; modern artificial networks frequently use fixed gradient descent rules and externally engineered buffers. Hassabis and colleagues describe reciprocal contributions between neuroscience and artificial intelligence, but reciprocal inspiration itself does not establish that the two kinds of system learn identically Hassabis et al. (2017). This distinction matters in a psychology of artificial systems: the appropriate evidence is the observed behavior and tested implementation of the configured artificial learner, with biological accounts providing constrained comparisons.

Replay, Reinstatement, and Generative Memory

Experience replay preserves selected earlier observations or representations and reintroduces them during later learning. It directly supplies data about old tasks that an optimizer would otherwise lack. Selection policies include random sampling, class balancing, diversity-oriented storage, and prioritization of informative cases; these decisions alter both the empirical performance and what the buffer represents. In small-memory settings, the essential experiment varies buffer size under matched memory accounting and reports whether current-task optimization is balanced against old examples. The study by Chaudhry and collaborators shows why even tiny episodic memories can offer competitive baselines, undermining claims that complexity itself proves superior scientific performance Chaudhry et al. (2019).

Deep Generative Replay substitutes generated examples for some retained historical data. Shin and collaborators proposed a generator that models previous training distributions and supplies synthetic earlier observations while a solver learns new tasks. The approach reduces reliance on storing every original record but shifts the problem to generator quality, coverage, and its own vulnerability to sequential change. A generator may reproduce common patterns accurately while losing infrequent but important cases; replaying its outputs can then reinforce an impoverished approximation of past experience. Evaluation therefore requires assessment of both downstream task retention and the representational adequacy of synthetic samples Shin et al. (2017).

Brain-inspired replay methods extend the hypothesis that reactivation can protect learned functions while preserving some flexibility for new experience. van de Ven, Siegelmann, and Tolias tested replay-based strategies in neural networks and presented a biologically motivated account of when generated or internal representations can sustain earlier performance van de Ven et al. (2020). Their results establish specific performance properties under the investigated task sequences. They also sharpen methodological questions about how replay and regularization interact, whether old input reconstruction is necessary, and what costs are incurred. Claims about biological consolidation and artificial rehearsal should remain separately attributed, with each method assessed by its own direct measurements.

Parameter Importance and Synaptic Regularization

Regularization approaches preserve earlier functions by penalizing certain parameter changes during later training. Elastic Weight Consolidation, introduced by Kirkpatrick and colleagues, estimates which model parameters are important for previous tasks and adds a quadratic penalty for departing from their earlier values. Its biological inspiration comes from ideas of synaptic consolidation, but its operational mechanism is a mathematical approximation around learned parameter estimates. The landmark 2017 experiments showed reductions in forgetting in supervised and reinforcement-learning settings. The evidence is strongest for the reported sequences and hyperparameter settings, not for unrestricted class growth or indefinitely changing real environments Kirkpatrick et al. (2017).

Synaptic Intelligence, presented by Zenke, Poole, and Ganguli, accumulates an importance signal from parameters' contributions along learning trajectories. Memory Aware Synapses, by Aljundi and colleagues, estimates importance through the sensitivity of network outputs to parameter changes, including under unlabeled observations. These methods differ in what they treat as evidence of importance, when estimates are updated, and how they penalize changes. Their common design is selective stabilization: parameters associated with retained competence receive greater protection. Across methods, a key vulnerability arises when different tasks require substantial changes to the same important parameters, making protection compete directly with acquisition Zenke et al. (2017) Aljundi et al. (2018).

A regularization study should show how parameter importance was calculated, how constraints accumulate over stages, and how hyperparameters were selected without privileged access to future or test data. It should measure whether retention gains survive longer sequences, varying domain shifts, limited model capacity, and class-incremental inference. Reports that omit new-task acquisition can conceal intransigence; reports that use task labels can conceal inference difficulty. Cross-scenario comparisons demonstrate that a regularization technique achieving strong task-incremental results may deteriorate in demanding class-incremental evaluations van de Ven et al. (2022). The scientific result is therefore a conditional mechanism with traceable assumptions, not a universal cure.

Architectural Isolation and Growing Capacity

Architectural approaches alter the pattern of parameter sharing instead of trying to protect every old function through a single penalty. Progressive Neural Networks allocate new columns for incoming tasks, connect them to earlier learned features, and freeze previously trained columns. This design preserves earlier parameterized solutions while allowing new modules to incorporate useful previous representations. It demonstrates a structural way to prevent a particular type of catastrophic overwriting, particularly when the inference mechanism knows which task-specific module to invoke. The cost is the growth of resources as more tasks are added, making parameter count, activation memory, and routing assumptions essential components of evaluation Rusu et al. (2016).

PackNet explores a different form of isolation: iterative pruning identifies parameters that can be freed after training a task, and the remaining task-relevant weights are protected as later tasks use newly available capacity. The work investigates how several visual recognition tasks can coexist within one network through masked parameter allocation. Such designs can substantially reduce direct destructive interference for tasks with stable assignments, but they introduce capacity-allocation questions: when should parameters be protected, how should scarce free capacity be distributed, and what happens when a past task needs to change? The treatment of stored masks and task identification determines whether the deployed solution remains resource efficient Mallya & Lazebnik (2018).

Structural separation changes the scientific meaning of retention. When earlier components are never altered, successful old-task testing can reflect preserved pathways rather than resilience of one continuously updated representation. This is a legitimate form of continual competence, but it addresses a different mechanistic problem than shared-weight consolidation. A complete experiment should distinguish performance of a fixed historical module from flexible integration of old and new information, and examine whether learned knowledge can be transferred across modules. Methods with expanding structures require long-horizon experiments that expose accumulated routing, storage, and energy costs rather than only the accuracy of the first few tasks Rusu et al. (2016) Wang et al. (2024).

Knowledge Distillation and Representation Preservation

Knowledge distillation constrains a new model by encouraging its outputs or internal representations to resemble an earlier model on selected inputs. The prior model serves as a reference for behavior that might otherwise be altered by learning newly introduced material. Distillation therefore retains information indirectly through response relationships rather than explicitly estimating the importance of every parameter. Its adequacy depends on what inputs are used to query the historical model and whether those inputs expose the boundaries of earlier decisions. If the new training set covers only novel categories, regularization against old outputs can still miss regions important for older classes Rebuffi et al. (2017) De Lange et al. (2022).

The iCaRL framework combines exemplar memory, representation learning, and a classification procedure for incremental classes. Rebuffi and colleagues evaluated a setting in which categories arrive sequentially while the learner must ultimately distinguish among all observed categories. Their framework illustrated how retained exemplars and distillation-style constraints can support a continually revised representation. The result should be interpreted as an integrated system: exemplar selection, feature adaptation, and classifier operation contribute together, so an improvement over a naïve sequential baseline cannot be credited to distillation alone without ablation Rebuffi et al. (2017).

Preserving outputs is a partial operational definition of preserving knowledge. Two models can give similar responses on a sampled retention set while their internal features diverge, and a high agreement rate can coexist with poor performance on unsampled inputs. Conversely, representational movement may be useful when it enables broader generalization. Scientific assessment should examine calibration, old-class decision margins, out-of-distribution cases, and stability under changes in the probing set. Claims that a representation has been retained require evidence about which functional relationships survived, not merely a low distillation loss during training Rebuffi et al. (2017) Wang et al. (2024).

Exemplar Selection and Class-Incremental Recognition

Class-incremental learning reveals forgetting particularly clearly because an expanding output space creates direct competition between old and newly introduced labels. A learner may correctly recognize a newly taught category when tested against only other new categories yet misclassify examples when every historical category is a possible answer. This is why protocols that reveal task identity at inference are inadequate substitutes for class-incremental tests. Exemplar rehearsal attempts to preserve old evidence under a bounded memory budget, and the choice of which examples to keep determines which aspects of the historical distribution remain represented Rebuffi et al. (2017) van de Ven et al. (2022).

GDumb is an instructive methodological challenge to elaborate sequential-update systems. Prabhu, Torr, and Dokania showed that maintaining a constrained memory of previous examples and training a model from scratch on that memory at evaluation can provide a strong baseline. This strategy changes the locus of complexity: incremental work concentrates on memory selection, while representation learning is postponed to evaluation. It may be unsuitable for continuously available low-latency inference, yet it forces comparisons to account for evaluation-time computation and the actual benefit added by elaborate mechanisms. Such baselines are scientifically valuable because they test whether claimed advances exceed straightforward data-management alternatives Prabhu et al. (2020).

Buffer construction may favor exemplars near feature centroids, boundary cases, diversity, or balanced representation of classes. Each selection strategy introduces assumptions about what the retained sample must accomplish. A centroid-oriented buffer may miss rare subpopulations; a boundary-oriented buffer may be overly sensitive to noisy or unstable labels; random selection may underrepresent infrequent conditions. Evaluation should report the number of stored images or tokens, feature representations used to choose them, class allocation rules, and whether memory grows with the number of classes. The central question is which evidence must remain available for a particular functional competence to survive later updating Chaudhry et al. (2019) Rebuffi et al. (2017).

Prompt-Based Continual Adaptation

Prompt-based continual learning can update a limited set of learned prompts or routing structures while retaining most parameters of a pretrained backbone. Learning to Prompt, introduced by Wang and colleagues, investigated the use of a prompt pool and input-dependent selection to support class-incremental visual recognition. The method changes the representation of the task-specific learning problem: adaptation can occur through selecting and optimizing prompts rather than revising every shared parameter. This may reduce interference in the pretrained backbone, but prompt capacity, retrieval quality, and the distribution on which prompt selection was trained become new determinants of stability Wang et al. (2022).

A frozen backbone does not guarantee a stable deployed function. The learned prompt pool can change, the retrieval mechanism can select different prompts for similar inputs, and the incoming distribution can make historical prompts inaccessible. Evaluation therefore needs both the final prediction and an account of which prompt or component produced it. Prompt isolation and routing can also require task-specific clues that are unavailable in open-ended usage. These operational details distinguish mechanisms that retain historical task skills from mechanisms that merely avoid weight modification. Taxonomies spanning parameter isolation, replay, and regularization clarify why prompt learning is one member of a larger strategy space Wang et al. (2022) De Lange et al. (2022).

In language-model settings, prompt instructions, retrieval context, parameter-efficient tuning, and full-model updates all produce behavioral changes, but their persistence mechanisms differ. Adding information to a context window provides access during that interaction; it does not by itself show that the model's weights were updated or that the information will survive a later session. A persistent external memory system can preserve records without the model learning them internally. Reviews of lifelong language-model adaptation emphasize the need to define which state is updated and how knowledge is evaluated under sequential tasks Shi et al. (2025) Zheng et al. (2025).

Transfer, Generalization, and Task Relationships

Transfer evaluates how experience with one distribution affects competence on another. Forward transfer occurs when earlier learning improves acquisition or evaluation on later tasks; backward transfer occurs when later learning improves earlier-task performance. Positive transfer and harmful interference can coexist because different features or examples participate in the two effects. Gradient Episodic Memory was motivated partly by the possibility of allowing updates that help earlier tasks while constraining those that would harm stored examples. Such procedures make task relationships experimentally visible rather than treating retention as a requirement to freeze an entire old solution Lopez-Paz & Ranzato (2017).

Generalization requires evaluation beyond the exact items retained in a replay buffer or the narrow domains on which a method was tuned. A learner that retains memorized exemplars might appear robust on a duplicated test but perform poorly when input conditions change. Conversely, updating shared features can support old tasks under new lighting, vocabulary, or context even while altering particular historic parameters. Good experiments separate retention on original held-out distributions from transfer to shifted or recombined distributions, and they report both improvements and failures. Surveys of classification-oriented continual learning describe benchmark diversity as essential because no finite sequence captures all modes of distribution change De Lange et al. (2022) Wang et al. (2024).

Task relatedness is dynamic. An initially unfamiliar domain can become easier after experience with intervening tasks, while a class distinction that once mattered can become invalid when the operational environment changes. Retention targets should reflect this history: knowledge that remains true should be maintained, and outdated mappings may need explicit revision. Evaluators can introduce recurring tasks, partially overlapping labels, and controlled shifts in causal structure to ask whether a model transfers principles rather than replaying isolated examples. This experimental agenda connects continual learning with the larger psychological problem of adaptive knowledge use across changing contexts van de Ven et al. (2022) McClelland et al. (1995).

Experimental Paradigms and Benchmark Design

A continual-learning benchmark is an experimental design specifying training sequence, input distributions, target labels or rewards, available memories, optimizer resets, and evaluation schedule. These details determine the scientific question being asked. Permuted-image tasks examine whether a shared learner can acquire transformed mappings, split-class protocols examine category growth, and domain shifts examine robustness when the labels remain comparable. Each isolates selected dimensions of sequential adaptation. Hsu and colleagues warned that comparisons across mismatched scenarios can make algorithmic rankings misleading, and van de Ven and colleagues demonstrated the distinct demands imposed by task-, domain-, and class-incremental settings Hsu et al. (2018) van de Ven et al. (2022).

Benchmark validity begins with a precise reference condition. Joint training on all tasks provides an informative upper comparison under privileged data access, while isolated single-task training indicates each problem's learnability. Plain sequential fine-tuning often supplies a lower retention baseline but can excel in current-task adaptation. Rehearsal with a fixed buffer and training from scratch with that same buffer test what complex mechanisms add beyond stored data. Performance differences should be interpreted together with differences in optimization passes, total data exposures, parameter capacity, and inference-time access to task identity. The challenge to earlier results offered by GDumb is a reminder that a strong comparator can change the interpretation of a research field Prabhu et al. (2020) Chaudhry et al. (2019).

Sequence construction introduces another validity issue. A favorable ordering can reduce conflicts between categories; an adverse ordering can greatly increase them. Results from a single random seed or a single fixed class order can be unstable because both initialization and sequence affect the path through parameter space. More informative designs sample multiple orders, run independent training seeds, and publish dispersion as well as central estimates. Recurrence, gradual drift, unknown boundaries, imbalanced classes, and delayed evaluation extend a benchmark beyond tidy task blocks. Such extensions are necessary when claims concern sustained deployment rather than performance on one controlled laboratory sequence Wang et al. (2024) De Lange et al. (2022).

Metrics, Confidence, and Inferential Validity

The performance matrix A_(t,i), recording test accuracy or another prespecified outcome on task i after learning stage t, is the elementary observational object in continual learning. Its rows show how one learner's capabilities change as experience accumulates; its columns show the trajectory of each historically introduced task. Average accuracy provides a compact end-state summary but conceals when and for whom losses occurred. A forgetting metric based on the gap between prior best and current performance explicitly quantifies deterioration, though it can be distorted by unstable earlier peaks. Backward transfer, forward transfer, and intransigence address related but distinct dimensions of the longitudinal process Lopez-Paz & Ranzato (2017) Chaudhry et al. (2018).

Accuracy should not be the only outcome when classes are imbalanced or confidence matters. Balanced accuracy, class-specific recall, log loss, calibration, and operational error costs may reveal harms hidden by aggregate scores. If an old class becomes infrequent, an apparently high global accuracy may coincide with substantial degradation on that class. Researchers should also report the stage at which a deficit first appears, the persistence of deficit across later stages, and whether the outcome recovers after re-exposure. A single final measurement cannot establish the time course of interference. For reinforcement-learning systems, reward curves must be paired with controlled evaluation conditions because policy behavior can change as the environment changes Wang et al. (2024) Kirkpatrick et al. (2017).

Statistical inference requires repeated independent runs and explicitly declared experimental units. Tasks within a sequence are not automatically independent replicates, and adjacent checkpoints from one trained model are repeated observations rather than separate experimental systems. Confidence intervals should be aligned with the relevant source of variation, whether seed, task order, data sample, or environment. When comparing methods, identical memory accounting and matched tuning budgets matter as much as significance thresholds. Claims about a universal algorithmic advantage are not warranted by one benchmark family or an unreported sweep of hyperparameters. Reporting the full score matrix and reproducible configuration allows later investigators to test alternative summaries De Lange et al. (2022) Hsu et al. (2018).

Strong Baselines, Reproducibility, and Negative Results

Scientific progress depends on comparison with implementations that remove unnecessary complexity. In the study of tiny episodic memories, simple experience replay emerged as a competitive strategy despite elaborate proposals for continual learning. GDumb then showed that a model retrained only from a managed memory buffer can rival more sophisticated systems on selected incremental tasks. These results matter because they identify which part of an apparently successful system carries causal responsibility. A method claiming superior retention must demonstrate an advantage over matched replay, storage, retraining, and capacity baselines under the same evaluation protocol Chaudhry et al. (2019) Prabhu et al. (2020).

Dark Experience Replay uses knowledge distillation over replayed examples, illustrating how a relatively compact design can combine stored data and soft historical targets. Buzzega and colleagues presented it as a strong baseline for general continual learning. Its contribution should be evaluated not merely by end-state classification but also by its performance under the chosen memory and compute budget. A model that performs additional operations during retrieval or trains with an enlarged historical buffer has received resources that may explain part of the gain. Reproducible comparisons should publish initialization, data order, augmentation rules, optimization schedule, buffer policy, and test-time inference procedure Buzzega et al. (2020).

Optimization dynamics supply another source of misleading conclusions. Mirzadeh and collaborators investigated how learning-rate schedules, training regime, and weight-space properties affect forgetting. Seemingly fundamental advantages may diminish after simpler systems receive comparably careful optimization. Negative results are therefore epistemically useful: they can reveal that a proposed memory mechanism is redundant under some conditions or that gains depend on an evaluation shortcut. Independent replication across codebases and datasets is stronger than repeated configurations within a single study. A mature experimental report distinguishes stable mechanistic findings from results contingent on a training recipe Mirzadeh et al. (2020) Wang et al. (2024).

Online Streams and Bounded Resources

Online continual learning requires a model to adapt under a stream that may be seen only once and that may contain nonstationary data, uncertain task boundaries, and resource constraints. The learner cannot assume that an entire new task dataset will remain accessible for multiple training epochs. This setting increases the importance of rapid sample selection, short-memory rehearsal, and scheduling decisions. A technique effective after extensive task-specific optimization can be impractical when each event arrives once and the system must remain continuously available. Distinguishing online from offline incremental learning is therefore a necessary part of the research passport for every result Chaudhry et al. (2019) Wang et al. (2024).

Resource accounting should include retained raw examples, features, optimizer state, protected historical models, specialized heads, prompts, and generated replay. Comparing a method with a compact exemplar buffer to an expanding architecture without reporting total parameters mischaracterizes the retention tradeoff. Energy, wall-clock adaptation latency, inference complexity, and storage policy become relevant as sequences lengthen. At a fixed memory budget, buffer allocation across classes determines whose past data remain accessible. A method that stores all earlier examples solves a materially easier version of continual learning than a method that must discard them. The experimental restriction must therefore be enforced and audited Rusu et al. (2016) Prabhu et al. (2020).

Evaluation under a data stream should probe the effect of gaps between repetitions and the recurrence of previous conditions. A learner may appear to forget permanently after a single new segment but recover quickly when a recurring context returns. Another may retain coarse classification while losing rare decisions that were sparsely represented in memory. Measuring recovery time, adaptability to drift, and worst-case retention provides a more informative picture than only final mean performance. Real systems may also face data retention regulations, making the distinction between replay from original records and reconstruction from allowed summaries practically consequential. These issues should be addressed as experimental constraints rather than deferred as deployment details De Lange et al. (2022) Aljundi et al. (2018).

Mechanistic Failure Analysis and Competing Explanations

Catastrophic forgetting is often described as old information being erased, but the observed failure can result from several distinct alterations. Shared features may shift away from their earlier useful configuration; a classifier's decision boundary may become biased toward newly introduced classes; a retrieval policy may select the wrong component; or an evaluation protocol may withhold task labels previously required by the method. These mechanisms generate overlapping behavioral signatures. Causal interpretation requires interventions targeting the suspected component, such as restoring a historical readout while preserving current features or reinstating the correct task route Rebuffi et al. (2017) van de Ven et al. (2022).

An informative failure analysis separates degradation in representation from degradation in access. Investigators can freeze feature extractors and retrain only a small diagnostic probe for an old task, asking whether the representation still contains discriminative information. They can evaluate stored exemplars and entirely new held-out examples separately, asking whether rehearsal preserved broad competence or merely isolated cases. They can vary task identity at inference, asking whether the model actually detects contexts or relies on privileged routing. They can restore earlier checkpoints and replay selected examples to measure recoverability. Each probe supports a narrower claim than a declaration that knowledge has disappeared from the system French (1999) Hsu et al. (2018).

Alternative explanations also concern the environment. A historical task may cease to be well-defined because labels change, sensors are updated, or norms governing correct answers evolve. Lower performance against an unchanged historical gold standard can then reflect adaptation to revised validity conditions. For scientific clarity, stable historical retention tests should be separated from controlled concept-drift tests where earlier labels are explicitly revised. The normative question of which information ought to be forgotten is distinct from the empirical question of why a measured score declined. Continual-learning psychology benefits from preserving this separation while investigating how adaptive systems balance competing temporal demands Wang et al. (2024) McClelland et al. (1995).

Continual Learning in Reinforcement-Learning Agents

Reinforcement-learning agents acquire action policies through interactions shaped by rewards, transition dynamics, and exploration. Sequential changes may alter not only predictive mappings but also which states are visited and which actions remain accessible. Catastrophic forgetting can appear as failure to perform previously mastered behaviors after a policy has been adapted to a new environment. This creates additional measurement problems: a change in accumulated reward may reflect the modified environment, a newly selected strategy, or impaired historical competence. Controlled evaluation must restore comparable earlier conditions and distinguish retention of a skill from its current priority under the agent's objective Kirkpatrick et al. (2017) Rusu et al. (2016).

Elastic Weight Consolidation was investigated in reinforcement-learning settings as well as supervised learning, providing evidence that constraining changes to important parameters can support the retention of earlier policies in particular sequences. Progressive Neural Networks also address sequential policies through separate learned components and lateral transfer. These approaches intervene at different levels. Regularization permits a shared representation to change under selective penalties; progressive expansion allocates structural capacity to new tasks. A rigorous comparison must match cumulative data exposure and report whether task identity is known at evaluation. In an open environment without external task labels, routing itself becomes a learning problem Kirkpatrick et al. (2017) Rusu et al. (2016).

Agent evaluation should combine episodic return, success rate for defined subtasks, recovery after context recurrence, and vulnerability to catastrophic errors. Retention may be particularly important for safety-relevant procedures that appear infrequently in the training stream. New policy learning can also make older behavior intentionally obsolete after environmental rules change. Experimental claims should therefore state whether earlier policies remain valid, whether the same observations are used across stages, and whether the agent retains its prior action space. These constraints help isolate interference from shifts in goals, environment, or observation and connect the analysis to functional psychology of artificial systems Wang et al. (2024) De Lange et al. (2022).

Continual Pretraining and Language-Model Adaptation

Large language models confront sequential adaptation through additional pretraining, instruction tuning, domain specialization, alignment procedures, and updates to facts or skills. A model trained further on a narrow domain may change general language capabilities or previously established task behavior; another update may improve recent knowledge while changing how older prompts are answered. Surveys by Zheng and colleagues and Shi and colleagues classify continual learning for language models across knowledge acquisition, domain adaptation, alignment, and evaluation. Their synthesis shows why the relevant dependent variables extend beyond class accuracy to factual recall, linguistic competence, instruction following, and cross-domain generalization Zheng et al. (2025) Shi et al. (2025).

Language-model forgetting requires careful task specification because a changed answer is not automatically evidence of lost information. The prompting context, decoding settings, sampling variability, safety rules, retrieval documents, and system instructions can alter outputs without any weight update. Controlled studies should freeze these variables when assessing parameter-level forgetting, then manipulate them independently to identify their contribution. A longitudinal design can maintain evaluation suites for older capabilities, current fine-tuning objectives, and transfer tasks, adding probes for hallucination, calibration, and robustness to paraphrase. If evaluation items are contaminated by training data or selected after model inspection, retention estimates lose validity Shi et al. (2025) Wang et al. (2024).

Different interventions solve different persistence problems. Replay during continued training may preserve historical distributions; parameter-efficient adapters can isolate some updates; selective regularization can protect earlier functions; external retrieval can restore access to records without changing the underlying model's weights. A knowledge repository may be durable while a language model's use of that repository remains context-sensitive and unreliable. Empirical claims should therefore distinguish stable external storage, learned retrieval policies, and changes in the pretrained network. This decomposition is essential when translating continual-learning literature into an account of long-running artificial assistants Zheng et al. (2025) Wang et al. (2022).

Configured Artificial Systems and External Memory

A deployed artificial system may combine a base model, prompts, persistent files, search mechanisms, tools, retrieval indexes, task planners, and logged interaction history. The observable unit of psychological-style investigation is then the configured system acting in an environment across occasions. Internal parameter learning is only one possible source of behavioral change. A system can appear to retain an earlier conversation because an external record is retrieved; it can also fail to use that record despite accurate storage. A claim about continual competence must identify which component preserves the relevant information and under what conditions that information becomes behaviorally available Shi et al. (2025) Zheng et al. (2025).

External memory changes the stability–plasticity tradeoff by moving some information away from shared trainable parameters. Updating a retrieval index may make current facts easier to access without altering old parameterized mappings; the same update may make older records less likely to be selected. This creates a functional analogue of interference at the level of retrieval and routing, though the causal mechanism differs from catastrophic gradient-induced forgetting. Investigators can isolate the components by holding the model fixed while varying the index, holding the index fixed while changing prompts, and then changing model weights with everything else controlled. The resulting factorial design attributes behavior to identifiable mechanisms McClelland et al. (1995) Wang et al. (2024).

Artificial Era research can study the temporal organization of such systems without assuming that output continuity demonstrates human-like remembering or subjective experience. Useful constructs include longitudinal response stability, recovery of earlier task policies, retention of explicit records, calibration of claims about the past, and adaptation to corrected evidence. These constructs can be operationalized with repeated interactions and auditable retrieval traces. A system that responds consistently over time can still be using an external source, and a system that revises its responses may be correcting earlier mistakes. Scientific interpretation depends on the configuration, intervention, and independent evaluation of competence rather than on anthropomorphic labels Hassabis et al. (2017) van de Ven et al. (2022).

Selective Forgetting, Data Governance, and Safety

Responsible continual learning includes decisions about which material should remain accessible and which should be revised or removed. Retaining obsolete personal data, withdrawn permissions, or superseded instructions can produce concrete harm even when a benchmark credits perfect retention. Conversely, abrupt updates can disrupt safety-relevant behaviors that remain valid. Research should separate three operations: preserving earlier functional competence, revising information whose truth conditions have changed, and deleting data whose retention is impermissible. These objectives may require different mechanisms and cannot be inferred from one classification accuracy score Wang et al. (2024) De Lange et al. (2022).

Replay methods create specific governance obligations because historical samples can remain in buffers or generative models. If previously collected data must be deleted, an experiment needs to specify whether raw records, learned representations, derived features, or synthetic reproductions remain. A model may continue to produce traces of earlier examples even after source files are removed. The original purpose of generative replay was to preserve useful training distribution information, not to establish compliance with deletion requests. Evaluating selective forgetting consequently requires independent privacy and data-removal tests in addition to historical-task performance Shin et al. (2017) Chaudhry et al. (2019).

Continual adaptation can also affect reliability through unpredictable regressions. A system deployed in education, health information, or automated decision support may lose previously tested abilities after an update. The scientific response is predeclared retention suites, regression tests for high-consequence tasks, monitoring under distribution shift, documented rollback procedures, and explicit decision authority for approving new model versions. These are engineering and governance implications derived from observed interference risks; they do not constitute proof of clinical effectiveness or regulatory compliance. Evidence from a laboratory benchmark should be translated to consequential applications only through domain-specific validation Aljundi et al. (2018) Shi et al. (2025).

Unresolved Scientific Questions and Research Agenda

A central unresolved question concerns what a scalable artificial learner must retain about its prior experience to preserve competence over an unbounded sequence. Fixed exemplar buffers sample the past but eventually confront representation and coverage limits; expanding architectures avoid some direct overwriting while increasing resource demands; regularization can protect learned functions while constraining future learning. Each strategy offers a measurable hypothesis about the sufficiency of its memory representation. Long-duration experiments with controlled shifts and recurring conditions can determine which information is necessary for retention and which historical detail can safely be discarded Wang et al. (2024) De Lange et al. (2022).

Another question concerns the explanation of forgetting. Correlations between weight changes and earlier test errors are weaker evidence than targeted interventions that restore or disrupt selected components. Researchers should compare old-model feature probes, historical task routes, memory-buffer substitutions, and carefully matched optimizer resets in the same experimental setting. They should report recovery dynamics rather than treating the first failure as permanent loss. Such experiments can test whether apparent forgetting is primarily representational degradation, biased readout, retrieval failure, or changed task validity. Theoretical interpretations remain provisional until differentiated by evidence that could favor an alternative account Ratcliff (1990) Mirzadeh et al. (2020).

Open-ended artificial systems require additional work on task boundaries, continual adaptation under uncertainty, and longitudinal measures beyond static datasets. A strong future protocol would maintain reproducible histories of data exposure, component versions, parameter changes, external-memory updates, and environmental shifts. Its outcomes would include historical retention, acquisition, transfer, calibration, worst-case errors, resource cost, and controlled forgetting of obsolete information. Replication across independent training pipelines, task orders, and populations of prompts would distinguish robust advances from convenient benchmark configurations. These requirements follow from the methodological evidence assembled here rather than announcing a newly verified theory van de Ven et al. (2022) Zheng et al. (2025).

Conclusion

Continual learning investigates how adaptive systems acquire new competence while retaining earlier useful capacities across changing experiences. Catastrophic forgetting is a family of observable sequential-learning failures in which subsequent learning substantially degrades previously demonstrated performance. The classical connectionist experiments established the vulnerability of shared distributed representations, and later computational work translated the problem into measurable task sequences, performance matrices, and controlled interventions. Cognitive memory research provides complementary models of interference and consolidation, while mechanistic conclusions about artificial networks must be grounded in their own training and inference operations McCloskey & Cohen (1989) McClelland et al. (1995) French (1999).

The evidence supports several conditional approaches to reducing forgetting. Replay reintroduces information from the past; importance-weighted regularization selectively protects parameters; architectural methods reserve capacity or isolate tasks; distillation preserves selected response relationships; prompt-based methods concentrate adaptation in specialized components. Their value depends on what information is available, which evaluation scenario is used, how much new learning is achieved, and what resources are consumed. Strong baselines and independent comparisons show that apparent superiority can change under matched storage, computation, and task definitions. The field therefore requires simultaneous measurement of stability, plasticity, transfer, and implementation cost Kirkpatrick et al. (2017) Prabhu et al. (2020) van de Ven et al. (2022).

The research agenda for the Artificial Era centers on identifiable, versioned artificial systems whose capabilities change through model updates, memory retrieval, environmental interaction, and revised policies. Longitudinal scientific evaluation must distinguish these mechanisms and determine which earlier functions should persist, which should adapt, and which information should be deliberately removed. Continual learning remains an experimental research program with important open questions about scalability, mechanism, safety, and the validity of real-world generalization. Its most durable conclusion is methodological: persistent artificial competence can be established through repeated controlled observation across an explicit history of learning and change Wang et al. (2024) Shi et al. (2025).

References