Catastrophic forgetting is one of those problems in AI research where the theory is better than the practice by an uncomfortably large margin. The mechanism is well understood: neural networks encode knowledge in the weights of the network, and when those weights are updated to accommodate new information, the updates overwrite prior information in ways that can be sudden and total. The understanding of why this happens — the representational overlap between old and new tasks, the absence of any equivalent to biological consolidation mechanisms — is fairly mature. What is less mature is the engineering that would make continual learning work reliably in the settings where it actually matters.
This gap between theory and practice has been persistent enough that it deserves examination. Why is the problem hard to close, and what would closing it look like?
The standard framing and its assumptions
Continual learning is typically framed as a sequence-of-tasks problem: a model trained on task A is then trained on task B, and we measure whether performance on task A degrades. In this framing, the goal is to reach the end of the task sequence with high performance on all tasks seen so far. The benchmark metrics — final accuracy on each task, backward transfer (how much learning task B affects performance on task A), forward transfer (how much learning task A helps with task B) — are well-defined and measurable.
This framing has produced a large body of useful work and has driven real methodological progress. But it has two assumptions baked in that, in many practical settings, are not met.
First, the tasks are assumed to be discrete and identifiable. The model is told, in most benchmark settings, when one task ends and another begins. In real deployment, tasks are typically not discrete — they shift gradually, and the model does not receive a signal that says "you are now in a different distribution". The task-boundary signal is a significant aid that real systems cannot rely on.
Second, the set of tasks is assumed to be finite, pre-specified, or at least well-characterised. Real knowledge accumulation is more open-ended: the relevant distributions change in ways that are not catalogued in advance, and the notion of "task" is itself a prior that may not fit the actual structure of the data stream.
Neither of these observations invalidates the benchmark framing — benchmarks have to make assumptions to be well-defined. But they flag where the translation from benchmark performance to practical capability is likely to be lossy.
The main families of approach
The literature on continual learning has converged on three main families of approach, each corresponding to a different intuition about what should be preserved when learning new information.
Regularisation-based methods impose constraints on how much the weights can change during new task learning, typically by penalising updates to weights that were important for prior tasks. The key challenge is estimating which weights are important, which requires maintaining some measure of parameter importance — an approximation to the Fisher information matrix being the canonical choice. Regularisation methods are computationally lightweight and do not require storing old data, but they scale poorly as the number of tasks grows, because the importance constraints from all prior tasks accumulate and eventually prevent any meaningful update to the network.
Rehearsal-based methods maintain a buffer of examples from prior tasks and interleave them with new training data. Catastrophic forgetting is reduced because the network continues to see examples from old tasks during new task training, preventing complete overwriting. The limitation is the memory requirement — storing raw examples is expensive — and the selection question: which examples should the buffer contain? Uniform random selection performs surprisingly well but is suboptimal, and smarter selection methods are an active area. Generative replay, which maintains a generative model of old tasks rather than raw examples, removes the raw storage cost but transfers the problem to model fidelity: the replay is only as good as the generative model.
Architecture-based methods allocate separate parameters or separate subnetworks to different tasks, preventing interference by structural isolation. Progressive neural networks are the purest form: each new task gets a new network column, with lateral connections to prior columns for transfer, and the prior columns are frozen. This provides complete protection against forgetting — prior weights are literally not updated — at the cost of growing model size with each new task, which is unsustainable at scale.
Where the practical failures occur
Having worked with all three families in our research, we have found that the practical failures tend to cluster around several specific issues that are underrepresented in the benchmark literature.
The first is implicit task shift without boundaries. When the training data distribution drifts continuously — the scenario most common in real data streams — all three families of approach perform worse than their benchmark numbers suggest. Regularisation methods miscalibrate their importance estimates because there is no discrete moment when importance should be reassessed. Rehearsal buffers fill with examples from early in the distribution rather than the current distribution. Architecture methods do not know when to allocate new capacity.
The second is evaluation calibration. Models trained with continual learning objectives are frequently overconfident on old tasks (the training signal reinforces old-task decisions long past the point of necessity) while being underconfident on new tasks (insufficient training before the evaluation point). These calibration issues matter for deployment in a way that accuracy-only benchmarks do not capture.
The third is interaction with the unlabelled-data problem. Most continual learning research assumes all training data is labelled. In our setting, where we are interested in systems that can accumulate knowledge from unlabelled data streams, the combination of continual learning and self-supervised objectives creates new interference dynamics that are not well-characterised. The contrastive objectives used in self-supervised learning are particularly sensitive to the composition of the current training batch — including rehearsal examples — in ways that can create subtle but persistent knowledge degradation.
Promising directions
The directions that look most promising to us are not the ones that most optimise the standard benchmarks. They are the ones that address the assumptions that the benchmarks make.
Online core-set selection — maintaining a compact, maximally informative replay buffer that is updated continuously rather than filled once — seems underexplored relative to its potential. The right formalisation probably involves something like a continual version of coreset approximation, where the buffer tracks the geometry of the seen distribution rather than specific examples.
Modular architectures with dynamic routing, where new inputs are directed to the most relevant existing module and occasionally trigger the creation of new modules when no existing module is a good fit, preserve structural isolation without requiring growth proportional to task count. The routing mechanism needs to work in task-boundary-free settings, which is the harder case, but it is the right structure for practical open-ended knowledge accumulation.
Finally, there is value in being clearer about what failure mode matters most in a given application. Not all forgetting is equally harmful. A system that forgets rarely-encountered edge cases while retaining the core distribution of its prior knowledge may be entirely acceptable for many applications. The research community's focus on zero-forgetting as the ideal has probably slowed progress on the more achievable and practically relevant goal of bounded, controlled forgetting — accepting some regression in exchange for simpler, more scalable implementation. The right trade-off depends on the deployment context, and that context is rarely specified precisely enough in the benchmark literature to guide the choice.
