Machines That Learn for Life: The Fundamental Problems of Continual Learning and the Most Promising Directions

A public-facing research synthesis · Conclusions first, argument and data after, all references in SOURCES.md. Chinese primary version: STUDY.zh.md.

For every reader, before the technical part begins: Today's AI learns the way a statue is cast, not the way a person grows — it is taught once on a huge pile of material, then frozen and put to use. If you try to teach it something new afterwards, the new knowledge tends to overwrite the old, because everything the system knows is stored in one shared set of internal numbers, and adjusting those numbers to learn one thing risks erasing another. Humans are the opposite: we learn all our lives, mostly keep our old skills, and the more we know, the faster we pick up the next thing. If this gap were closed, AI would stop being a product that must periodically be rebuilt from scratch and become more like a colleague who grows with experience — the longer you work together, the better it understands your world. This report explains why that is not possible today, where exactly the difficulty lies, and which research paths look most promising.

Evidence tags. Every key claim in this document is tagged by strength of evidence:

• [established] — replicated across independent studies, or rigorously proven; essentially uncontroversial in the field;
• [contested] — supported by published evidence, but with conflicting results or unresolved interpretive disputes;
• [speculative] — reasonable extrapolation or the authors' synthesis; no direct evidence yet; treat as hypothesis.

Only published/public literature is used (journals, conferences, arXiv preprints, public reports), cited as [S-numbers]; full bibliography in SOURCES.md. The original literature search ended in 2025; a 2026-09 update added a batch of individually verified public references (SOURCES.md section F, S100–S127). Venue names are stated only where directly verified.

A note on terminology. Every technical term is defined in plain words (in parentheses) the first time it appears. If you meet an unfamiliar word while skimming, its definition is at its first occurrence.

Part 0: Conclusions First

The motivating question: a truly strong AI should keep learning like a human — continuously, for its whole life. No such system exists today. What fundamental problems must be solved, and which directions look most promising?

Twelve core conclusions:

1. Today's deep learning is structurally two-phase — train first, then freeze and deploy — and this directly conflicts with what lifelong learning demands. It is not an engineering detail; it is a foundational defect. (A deep network is a learning system built from many layers of adjustable numbers; those numbers are its parameters or weights, and everything the model knows is stored in them.) The moment a standard deep network is made to learn continually (kept in use and taught as it goes, rather than trained once), it exhibits two failures at once: it forgets the old (catastrophic forgetting) and it stops being able to learn the new (loss of plasticity — plasticity being the capacity to still be changed by new experience). Both have solid experimental evidence. [established] [S1][S23]
2. Catastrophic forgetting has been known since 1989 and still has no general solution. Its mechanism is clear: in a distributed representation (each piece of knowledge spread across many shared numbers rather than filed in its own slot), old and new knowledge share parameters, and the gradient updates for a new task (the step-by-step adjustments made during training) overwrite the weights the old tasks depend on. The three mitigation families — regularization (penalizing changes so parameters cannot stray far from their old values), replay (keeping a small sample of old data and mixing it into new training as revision), and structural isolation (giving new tasks their own dedicated network parts) — each carry costs, and none holds up fully in the realistic setting of no task boundaries, single-pass data, and bounded memory. [established] [S1][S3][S13]
3. Loss of plasticity is the recently confirmed "other half" of the problem, and arguably the more lethal one. Dohare et al. (Nature, 2024) showed that standard deep-learning methods gradually lose the ability to learn in continual settings, eventually performing worse than a shallow network; gradient descent alone (the standard training algorithm: repeatedly nudging parameters in the direction that reduces error fastest) is not enough, and plasticity was maintained indefinitely only by algorithms that continually inject randomness — e.g., continual backpropagation, which keeps randomly resetting a small fraction of units that have stopped contributing, giving the network a supply of "fresh cells." [established as a phenomenon; the strong claim "randomness is necessary" is contested — see main text] [S23]
4. In theory, perfect continual learning "without memory, by clever parameter updates alone" is essentially hopeless. Knoblauch et al. (ICML 2020) proved that optimal continual learning generally requires perfect memory and is NP-hard (a class of problems that computer science regards as having no efficient solution); this explains why replay/memory methods systematically outperform pure regularization methods in practice. Any serious lifelong-learning design must include memory in some form. [established (the theorem); its practical force is reasoned inference] [S18]
5. Loss geometry gives "learning nearby" a real but bounded window of opportunity. (The loss is the score measuring how wrong the model is — training means pushing it down; "loss geometry" studies the landscape of that score over all possible parameter settings.) In overparameterized networks (far more parameters than strictly needed), good solutions are not isolated points but vast connected low-loss regions (mode connectivity); solutions obtained by fine-tuning (continuing training on new data) from the same pretrained model (a base first trained on huge general data) are often even connected by straight lines and can be directly merged by weight averaging (model soups, task arithmetic). This explains why the whole "stay near the previous solution" family — EWC, LoRA, orthogonal projections, model merging — works surprisingly well, especially on top of strongly pretrained models. [established] [S32][S33][S34][S40][S48][S49]
6. But "learning nearby" has principled limits: constraints accumulate and exhaust capacity, and genuinely new knowledge requires feature change, which cannot stay linearized forever. Orthogonal-projection methods (which only update parameters along directions that leave old tasks undisturbed) run out of usable directions as tasks accumulate; the Fisher estimates EWC relies on (a mathematical estimate of which parameters matter for old tasks) degrade over long sequences; pure regularization methods systematically fail in class-incremental settings (the hardest scenario, where the model is never told which task the current input belongs to); and learning in the lazy (NTK) regime (a state where parameters barely move and the model behaves almost linearly) is precisely learning without feature learning (the network inventing new internal concepts of its own) — giving up the thing that makes deep learning powerful. Learning nearby is a strong short-to-medium-term strategy on top of pretrained models, not a complete answer to lifelong learning. [established (each failure point) + synthesis] [S13][S37][S38][S46]
7. The brain's answer is not one mechanism but a multi-timescale systems architecture: complementary learning systems (fast hippocampus, slow consolidating neocortex), sleep replay, synapses with internal multi-timescale states, sparse coding, and neuromodulatory gating. (The hippocampus is the brain region that rapidly records specific events; the neocortex slowly extracts general patterns; a synapse is an adjustable-strength connection between neurons — the brain's "weight"; consolidation is the process of converting fragile new memories into durable ones; sparse coding means only a few neurons are active at any moment; neuromodulators are chemical signals, such as dopamine, that control when and how strongly to learn.) The most transferable insight: memory must be hierarchical, consolidation must be an active offline process, and "how fast to learn" must itself be a controlled signal. [established (biology); "transferable" is grounded inference] [S61][S86][S88][S89][S90]
8. The strongest engineering approach today, empirically, is the combination "memory system + replay + slow consolidation," not any single algorithm. Replay is the most robust method family across continual-learning benchmarks (standardized test suites); retrieval augmentation (RAG) — having the model look facts up in an external document store before answering, instead of relying only on what is baked into its weights — sidesteps weight updates entirely; and continual pretraining of LLMs (large language models) with "learning-rate re-warming (the learning rate is the step size of each parameter adjustment) + a small replay fraction of old data" already approaches full retraining performance. [established (individual results); the prioritization is the authors' judgment] [S13][S64][S74]
9. In-context learning (ICL) is a real but ephemeral form of continual learning, and it points to an architectural way out: make learning part of inference. (In-context learning: the model picks up a new task from examples placed in its context window, with no parameter changes at all; inference is the phase when a model is being used to produce outputs, as opposed to being trained.) A transformer's forward pass (the single computation that produces one output) can implicitly implement gradient descent — learning "in its head"; test-time training (TTT layers, Titans) makes this explicit, letting the model continually update a fast memory module while it runs. This is among the most imaginative current architectural directions, but it has not been validated at genuinely lifelong scale. [established (mechanistic studies)] / [speculative (as a lifelong-learning route)] [S67][S68][S69][S70]
10. The field's standard framing partly asks the wrong question. Mainstream benchmarks define continual learning as "approximate joint multitask training (pooling all tasks' data and training on everything at once — treated as the ideal ceiling) under constraints" — the goal is not to lose points. But the true goal of lifelong learning is knowledge accumulation with compounding returns: the more you have learned, the faster you learn the next thing (forward transfer — old knowledge helping new learning), and old knowledge is improved by new experience (backward transfer). Almost no benchmark directly measures "learning acceleration"; and a naive baseline (the simplest possible comparison method — here GDumb: just store samples and retrain from scratch before the test) beats a large fraction of sophisticated methods — evidence that current evaluations do not measure the capability we actually want. [established (GDumb et al.)] + [position (the framing critique)] [S15][S16][S17]
11. The lesson from neurogenesis: even biology is still arguing with itself. (Neurogenesis: the brain growing new neurons.) Whether the adult human hippocampus produces new neurons for life was answered in opposite ways by two top-venue papers in 2018 (Sorrells: undetectable in adults; Boldrini: persists to age 79). The implication for AI is directional (injecting fresh plastic units is valuable — exactly what continual backpropagation does), not a blueprint to copy. [contested (the biological fact)] [S23][S94][S95]
12. Ranked most-promising directions (authors' synthesis; reasoning in Part 6): ① complementary learning systems for LLMs (retrieval-based episodic memory — memory of specific experiences, as opposed to memory of general patterns — plus periodic consolidation into weights); ② plasticity-preserving optimizers and architectures (the continual-backpropagation family); ③ a "recipe science" of continual pretraining at scale; ④ architectures with learning built into inference (TTT / fast weights); ⑤ geometry-guided model merging and low-rank updates (freezing the model's body and training only a small attached set of parameters — LoRA is the best-known example). The real bottleneck is not any single technique but the absence of an honest open-ended lifelong-learning evaluation and a theoretical objective function. [speculative (the ranking)]

Part 1: Problem Definition — What Is Continual Learning Actually Asking For?

1.1 The prototype dilemma: stability vs. plasticity

A learning system facing an endless stream of experience can die in two symmetric ways: too stable, and nothing new gets in; too plastic, and the new overwrites the old. Grossberg named this the stability–plasticity dilemma in 1980 [S4] and made the first architectural attempt at it with ART [S5]. This is not a quirk of deep learning; it is a universal constraint on any system that stores knowledge in shared resources.

Deep networks exhibit the dilemma in extreme form: McCloskey and Cohen (1989) found that sequentially trained connectionist networks (an older name for neural networks) catastrophically destroy the first task's performance while learning a second [S1]; Ratcliff (1990) quantified it systematically [S2]. French (1999) diagnosed the root cause: the power of distributed representations comes precisely from parameter sharing and overlap, and overlap is interference — forgetting is not a bug; it is the flip side of the same coin as distributed representation [S3]. [established]

1.2 The standard formalization and its metrics

Modern papers usually set the problem up as a sequence of tasks T₁, T₂, …, Tₖ arriving one at a time, with the learner seeing only the current task's data, and the goal of performing well on everything seen so far. Lopez-Paz and Ranzato (2017) defined the three metrics still in use [S15]:

• Average accuracy (ACC) across all tasks after training ends;
• Backward transfer (BWT): the effect of new learning on old tasks (negative = forgetting; positive = old skills improved);
• Forward transfer (FWT): how much old knowledge helps tasks not yet seen.

van de Ven et al. further split the problem into three scenarios [S12][S13]: task-incremental (task identity given at test time), domain-incremental (same task, shifting distribution), and class-incremental (discriminate among all classes seen so far, without being told the task). The distinction matters enormously: regularization methods (EWC and kin) are passable in task-incremental settings but fail systematically in class-incremental ones, where essentially only replay and structural methods survive [S13]. [established]

1.3 Does this framing ask the right question? — Our argument

Partly right, partly wrong. Three criticisms, each backed by published evidence:

First, current benchmarks measure "not regressing," not "accumulating." The standard metrics take joint multitask training as an implicit ceiling: a model trained on all data pooled together is the optimum, and continual learning is defined as approximating it under a "no pooling" constraint. But the original motivation of lifelong learning (Ring 1994 [S6]; Thrun & Mitchell 1995 [S7]) was compounding knowledge: the more you learn, the faster and better you learn next — the learning curve on task 1,000 should be far steeper than on task 1. That is forward transfer, which in most papers is a secondary, in-passing metric; almost no work demonstrates learning acceleration growing with task count. A system that merely "doesn't forget" but never "learns faster" is not the lifelong learner anyone wants. [position, but the state of the metrics is verifiable] [S15]

Second, the victory of naive baselines exposes distorted evaluation. GDumb (Prabhu et al., 2020) uses no continual-learning technique at all: it greedily fills a small buffer with samples and retrains a model from scratch before testing — and beat a large swath of purpose-built methods [S17]. Farquhar and Gal (2018) had already shown that common protocols (multiple passes over data, test-time task labels, short two-or-three-task sequences) bias evaluation toward methods that do not address the real problem [S16]. A later re-evaluation under fixed compute budgets found that many published gains vanish once compute is matched, with simple methods (replay plus sensible allocation of compute) often winning [S124]. If a benchmark can be cracked by a method that does no continual learning, and gains do not survive budget matching, the benchmark is not measuring the target capability. [established]

Third, real lifelong learning has no task boundaries — and "task" itself may be the wrong abstraction. Real experience streams are continuous, non-stationary, and unsegmented. Task-free continual learning (Aljundi et al. 2019 [S19]), online continual learning (single-pass streams; empirical survey in Mai et al. [S122]), and general continual learning (Buzzega et al. 2020 [S55]) move in the right direction but remain a minority. Silver and Sutton's "Era of Experience" (2025) raises the bar further: agents should inhabit lifelong streams of experience, with goals, memory, and learning spanning the whole stream rather than being chopped into sessions [S20]. By that standard, nearly all continual-learning literature studies a toy projection of the real problem. [position; the cited stance is published]

In fairness to the standard framing: a toy projection is not useless. Catastrophic forgetting and plasticity loss already appear in the cleanest small settings, and the mechanistic understanding in Parts 2–3 was won precisely through such controlled setups. The criticism targets mistaking benchmark scores for progress on lifelong learning itself, not the methodology of controlled experiments.

1.4 A theoretical anchor: no free memory

Knoblauch et al. (ICML 2020) proved, in a set-theoretic framework, that an optimal continual learner that avoids catastrophic forgetting in the worst case must generally solve an NP-hard problem and requires perfect memory (retaining at least one element of every "equivalence set") [S18]. Two corollaries:

1. Hoping for a clever update rule that solves forgetting without storing any information about the past is theoretically unfounded;
2. This explains the empirical landscape: replay/memory methods systematically beat pure regularization [S13][S17].

So the question shifts from "whether to remember" to "what to remember, in what form (raw samples / generative model / weights / external store), when to consolidate, and when to discard." That is the through-line of Parts 4 and 6. [established (theorem); corollaries are grounded interpretation]

One objective-level subtlety deserves its own flag: "zero forgetting" is not always the right goal. When the objective itself drifts over time, forgetting outdated knowledge can be exactly right — in recommendation, for instance, stale preferences are not just useless but harmful once a user's interests shift, and PESO explicitly builds its design on that argument [S117]. This suggests distinguishing interference (new learning accidentally damaging still-valid old knowledge — bad) from revision (deliberately overwriting invalidated old knowledge — good); nearly all current metrics collapse both into a single "forgetting" number. [established (the published argument); "interference vs. revision should enter the metrics" is our position]

Part 2: The Two Failures — Catastrophic Forgetting and Loss of Plasticity

2.1 Mechanisms of catastrophic forgetting

Systematic deep-learning-era evidence begins with Goodfellow et al. (2013) [S24]. The clearest current mechanistic picture:

• Shared parameters get overwritten. New-task gradients do not "know" which weights carry old functions; large learning rates, long training, and representational overlap between tasks all worsen the overwriting [S3][S24].
• Forgetting concentrates in deeper layers. Ramasesh et al. (2021) found that drift of representations near the output is the dominant source of forgetting, with earlier features comparatively stable; inter-task semantic similarity affects forgetting non-monotonically [S29]. [established]
• Scale is a palliative, not a cure. Ramasesh et al. (2022) showed larger, better-pretrained models forget less when fine-tuned — big models tend to accommodate new tasks in more "orthogonal" ways [S30]; Mirzadeh et al. (2021) independently found that making networks wider substantially reduces forgetting (deeper does not) [S106]. Both are consistent with the geometry of Part 3. But "less" is not "none": forgetting in continually fine-tuned LLMs is repeatedly observed (Luo et al. 2023 [S76]). [established]
• Flatness correlates with retention. Converging to flatter minima (smaller decayed learning rates, appropriate regularization) yields less forgetting [S39]. [established (correlation); causality contested]

2.2 Loss of plasticity: the other half

Forgetting is about losing the old; plasticity loss is about becoming unable to learn the new. It was long invisible because the standard paradigm trains only once. The evidence chain has filled in fast:

• The warm-starting penalty. Ash and Adams (2020): continuing training from an old model on enlarged data generalizes worse than retraining from scratch — even when the data simply grew [S25].
• Primacy bias and dormant neurons in deep reinforcement learning (RL — the paradigm where an agent learns behavior by trial and error from reward signals). Early data "sets" the network and degrades later learning (Nikishin et al. 2022 [S26]); ReLU units (the most common type of artificial neuron) die or go dormant in growing numbers — outputting zero and no longer participating in learning — and periodically resetting them helps substantially (Sokar et al. 2023 [S27]).
• The decisive evidence — Dohare et al. (Nature 2024) [S23]: on Continual ImageNet (thousands of sequential binary classification tasks), slowly drifting regression (regression: predicting a continuous number), and non-stationary RL, standard backpropagation's ability to learn decays with task count until it is worse than a shallow network. Accompanying signatures: rising fractions of dead units, growing weight magnitudes, and collapsing effective rank of representations (i.e., the network ends up using ever fewer independent internal dimensions — its expressive capacity shrivels). Their key claim: within everything they tested, only algorithms that continually inject random diversity maintain plasticity indefinitely — their continual backpropagation randomly reinitializes a small fraction of low-utility units and sustains learning in both supervised and RL settings. [established (the phenomenon and their in-scope conclusion)]

Worth flagging separately is the strong corollary — "gradient-descent methods are not enough; sustained deep learning requires a random, non-gradient component" [S23]. That is a universal claim, while the evidence is existential (every deterministic method they tried failed). Deterministic methods such as L2 Init (regularizing toward the initial parameters rather than the origin [S31]) also maintain plasticity under milder non-stationarity, so the boundary is not yet drawn. Our tags: the phenomenon [established]; the mechanistic account (curvature loss, rank collapse, dead units [S27][S28]) [established but incomplete]; "randomness is necessary" [contested].

2.3 Two failures, one coin

Put both back into Grossberg's dilemma [S4]: forgetting = insufficient stability; plasticity loss = the accumulated price of the stability the system buys itself (growing weights, ossifying representations). Any method that solves only one half pays on the other: freezing parameters to protect old knowledge accelerates plasticity exhaustion; randomly resetting units to restore plasticity locally destroys old knowledge. The core of lifelong learning is not to eliminate either failure but to find mechanisms that hold both axes simultaneously over long horizons — which is exactly what the brain's multi-timescale architecture does (Part 5). [synthesis]

Part 3: The Geometry Angle — How Far Does "Learning Nearby" Go?

3.1 Low-loss regions are far larger than intuition suggests

A counterintuitive fact about overparameterized networks: good solutions are not islands but a connected continent.

• Mode connectivity. Garipov et al. and Draxler et al. (2018, independently and simultaneously) found that minima (parameter settings where the loss is lowest) from independent training runs are joined by simple curves along which the loss barely rises [S32][S33]. [established]
• Linear mode connectivity (LMC). Frankle et al. (2020): if two solutions share an early-training starting point, they are often connected by a straight line of low loss; networks "stabilize" into a basin (a low-lying valley of the loss landscape) early in training [S34]. [established]
• Possibly one basin modulo permutations. Entezari et al. (2022) conjectured that after accounting for neuron-permutation symmetry (swapping two neurons along with their wiring leaves the network's function unchanged, so every solution has an enormous number of "renamed" twins), the solutions of SGD (stochastic gradient descent, the standard training algorithm) essentially land in a single linearly connected basin; Git Re-Basin (2023) empirically supported much of this via permutation alignment [S35][S36]. [contested (the conjecture's scope)]

The most directly relevant result for continual learning is Mirzadeh et al. (2021): given a shared initialization, the joint multitask solution is linearly connected to the continual-learning solution — a good non-forgetting solution lies "next door" to the continual learner's trajectory, and forgetting is the optimizer wandering out of that linear corridor for lack of a constraint [S40]. [established]

Two further characterizations round out the "low-loss continent" picture: large-scale analyses of high-dimensional loss landscapes show low-loss regions extending as broad wedge-like structures [S110], and "intrinsic dimension" experiments show many tasks can be solved inside a subspace far smaller than the parameter count [S111] — this surplus of solution space is the capacity behind "a good solution is usually nearby." [established (each experiment); the connection to continual learning is our synthesis]

3.2 Lazy training: why "next door" is often enough

NTK theory (Jacot et al. 2018 [S37]) and the lazy-training analysis (Chizat et al. 2019 [S38]) show that sufficiently wide networks barely move in parameter space during training; function change is approximately that of a model linearized at initialization. Width plus strong pretraining pushes models toward a "locally linear, locally solvable" state: a new task can often be solved within a linear subspace near the old solution. This is the unified theoretical intuition behind the entire learning-nearby family. [established (theory); its applicability to real pretrained models is an approximation]

Within this framework, forgetting itself already has a computable theory: Doan et al. (2020) quantify inter-task interference via the NTK overlap matrix — the overlap of old and new tasks in kernel-feature space [S107]; Bennani et al. (2020) give generalization guarantees for orthogonal gradient descent in the NTK regime [S108]; and Evron et al. (2022) characterize worst-case forgetting exactly in linear regression [S109]. The upshot of this theoretical thread: at least in the (approximately) linear regime, "which task combinations will damage each other" is in principle calculable in advance — connecting directly to P7 in Part 6. [established (each result, under linear/NTK assumptions)]

3.3 The "learning nearby" family, organized by how "nearby" is defined

1. Soft constraints (anisotropic springs). The plainest form is L2-SP — a straight L2 penalty toward the starting point (the pretrained parameters), long established in transfer learning [S101]; EWC upgrades that spring to an anisotropic, Fisher-information-weighted version — important parameters tethered tightly, unimportant ones loosely [S41]; SI accumulates importance online [S42]; MAS estimates importance without labels (i.e., unsupervised) [S43].
2. Function-space "nearby." LwF constrains not the parameters but the behavior — distillation makes the new model keep imitating the old model's outputs on new data [S100]. Parameter distance and function distance are not equivalent (see §3.4), so this is a substantively different definition of "nearby."
3. Hard constraints (orthogonal subspaces). OGD projects new gradients onto the orthogonal complement of old-task gradients [S44]; OWM uses orthogonal projectors plus a context module and demonstrated hundreds of sequential tasks in Nature Machine Intelligence [S45]; GPM explicitly maintains the core gradient subspace of old tasks [S46]; O-LoRA carries the same idea to LLM adapters, keeping each task's LoRA subspace orthogonal to the others [S114].
4. Low-dimensional increments. LoRA freezes the backbone and trains only a low-rank correction [S47] — "nearby" parameterized as a low-rank manifold; naturally supports one adapter per new domain. Large-scale evidence confirms both sides of the bargain directly: LoRA learns less and forgets less (Biderman et al. 2024 [S113]) — restricting the update's degrees of freedom buys retention at a measured cost.
5. Weight-space arithmetic. Task vectors (fine-tuned minus pretrained weights) can be added to compose skills and subtracted to unlearn [S48]; multiple fine-tunes of one starting point can simply be averaged (model soups [S49]) or merged with Fisher-information weighting [S116]; TIES additionally locates and resolves parameter interference during merging (sign conflicts, redundancy) [S115]; interpolating between fine-tuned and original weights retains both old and new performance (WiSE-FT [S50]). All of this holds only within a shared pretrained starting point (one linear basin).
6. Continualized merging and anchoring (recent developments). The merging above is mostly one-shot; a newer generation makes it a rolling process — SLAO merges each new task's update into a single LoRA via orthogonal initialization plus time-aware scaling, keeping memory constant in the number of tasks (ICLR 2026 [S118]); PESO maintains one continually evolving LoRA anchored by a proximal term to its most recent frozen state, and explicitly argues that in its recommendation setting outdated knowledge should be overwritten (ICLR 2026 [S117]).
7. "Nearby" in reverse. L2 Init tethers parameters toward the initialization rather than the old solution — the goal being plasticity rather than retention [S31]. The same geometric tool, used in both directions, maps exactly onto the two ends of the dilemma.

As an aside, "bound the size of each update" is a principle broader than continual learning: trust-region methods in reinforcement learning embody the same idea, and a recent public preprint on gradient-free optimization of recurrent spiking networks (arXiv:2601.21572 [S127]) stabilizes training with a trust region scaled adaptively by signal energy — an RL/optimization work only indirectly related to continual learning, noted here purely for the shared principle.

3.4 Where it breaks

Failure points, all [established]:

• Capacity exhaustion. Every task consumes a slice of subspace under orthogonal projection; usable dimensions shrink monotonically, so with enough tasks the method must either interfere or stop learning. This is arithmetic necessity, discussed in the GPM paper itself [S46].
• Drifting importance estimates. EWC's Fisher is a local quadratic approximation at the old solution; it degrades with distance, and accumulated penalties over-constrain long sequences (analyzed in Progress & Compress [S57]).
• Scenario mismatch. In class-incremental settings, regularization methods as a class come close to failing outright — they can preserve representations but not the decision boundaries between old and new classes, which require old and new data present together (or replayed) to calibrate [S13].
• Merging requires one basin. Task arithmetic and soups fail across different initializations; permutation alignment recovers only part of the effect [S35][S36][S48].
• Loss-nearby ≠ mechanism-nearby. Work on "mechanistic mode connectivity" shows that two solutions joined by a low-loss path can implement entirely different internal computations (e.g., relying on different features) [S112]; Ramasesh et al. likewise show deep-layer representations can drift substantially even when parameters move little [S29]. So "parameters staying nearby" does not guarantee "behavior on old data staying nearby" — precisely the divide between function-space constraints like LwF [S100] and parameter-space constraints. Which metric "nearby" should even be measured in — isotropic L2 [S101], Fisher-weighted [S41], or function-space distance [S100] — is itself an open question.
• Most fundamentally: lazy-regime learning = no feature learning [S38]. Genuinely new knowledge — new representational primitives, not recombinations of old features — requires leaving the linearized neighborhood. A system that stays near the pretrained solution forever has, as its ceiling, "remixes of pretrained features."

Verdict: geometry explains both why learning nearby works so surprisingly well on strong pretrained models (the low-loss continent is wide; locally, everything is nearly linear) and why it cannot suffice (every continent has a coastline; linearization cannot grow new features). It is an essential component — especially the engineering path of modular, mergeable updates around a strong base model — but not the full answer to lifelong learning. [synthesis; supporting points established]

📦 Aside: our toy experiment (illustrative only)

To make this part's claims tangible, we ran a deliberately tiny controlled experiment (via a companion reproducible-experiment repository). It is an illustrative demo, not evidence on par with the literature above: the dataset is tiny, seeds are few, no statistical claims are made, and only qualitative directions are reported.

• Setup: a 5-task "permuted digits" stream (sklearn's 8×8 handwritten digits — a miniature MNIST-like set; each task applies a fixed pixel shuffle); a numpy multilayer perceptron at widths 32/128/512; methods: plain fine-tuning, an L2 penalty toward the previous task's weights (L2-to-previous), EWC, a small replay buffer, and updates restricted to rank 4. Metrics: parameter distance moved per task, forgetting, forward transfer, and the loss barrier along the linear interpolation between consecutive task solutions. Runs in ~20 seconds on a CPU.
• Qualitative findings: replay was strongest (highest final accuracy, lowest forgetting) — directionally matching the literature landscape of Part 4 [S13] and the memory-necessity theory [S18]; L2-to-previous and EWC both beat plain fine-tuning, with L2-to-previous moving parameters least among full-rank methods — consistent with the learning-nearby picture; wider networks generally raised final accuracy and reduced forgetting for fine-tuning, L2-to-previous, and replay — directionally consistent with the overparameterization / "good solutions nearby" account [S30][S106]; and rank-4 low-rank updates moved least but underfit — a miniature of the "restrict freedom, buy retention, pay in learning capacity" trade-off of §§3.3–3.4 (the same phenomenon [S113] documents at scale).
• Limits: toy dataset; highly homogeneous tasks (permutation tasks barely conflict semantically, dodging the hard class-incremental case); few seeds. Its entire value is showing that Part 3's qualitative predictions are visible in a minimal setting reproducible in 20 seconds — not providing new evidence for or against any method.

Part 4: Systems Approaches — Replay, Dual Systems, External Memory, and the LLM Era

4.1 Replay: the least elegant, most effective

Rehearsal stores a small fraction of old data and mixes it into new training; the idea goes back to Robins (1995), who also proposed pseudorehearsal — reviewing with self-generated samples instead of stored data [S51]. The modern lineage: iCaRL (class-incremental with exemplar management) [S52], GEM/A-GEM (old-sample gradients as update constraints) [S15][S54], experience replay in reinforcement learning [S102], DER (replaying logits — the model's raw output scores, which carry more information than just the correct label; a very strong simple baseline) [S55], generative replay (a generative model synthesizes old data) [S53], and brain-inspired generative replay (replaying internal representations rather than raw pixels, with large class-incremental gains) [S103].

The empirical picture is remarkably consistent: in hard scenarios like class-incremental learning, replay is the most robust method family [S13][S14]. That is no coincidence — Knoblauch's theorem says memory is necessary [S18], and replay is memory in its plainest form. Generative replay escapes raw-data storage in principle but hands the problem to a recursion — the generator itself must learn continually — and has not yet truly worked on complex distributions. [established (the landscape); the judgment on generative replay is synthesis]

4.2 Complementary learning systems: the brain's architectural answer

The classic theory of McClelland, McNaughton, and O'Reilly (1995) [S61]: the brain resolves the dilemma with two learning systems — the hippocampus rapidly stores specific episodes in sparse codes, while the neocortex slowly learns statistical structure in overlapping distributed codes; during offline periods (sleep, rest) the hippocampus replays experience to the cortex, which absorbs it interleaved with old knowledge, avoiding interference. Kumaran et al. (2016) updated the theory and drew the explicit implication for AI agents: fast/slow dual systems + interleaved replay + episodic memory enabling one-shot learning [S62]. [established (theoretical standing and neural evidence; see Part 5)]

The engineering counterparts already exist, though rarely under this name:

• Structural isolation = giving the "fast system" its own capacity: Progressive Networks add a column per task [S56]; PackNet frees capacity by pruning (removing unimportant connections) [S58]; HAT learns per-task masks (switching different neurons on or off per task) [S59]; XdG uses sparse context gating so different tasks activate nearly non-overlapping units, and combined with synaptic stabilization holds ~90%+ accuracy across 500 tasks [S60]. The lightweight pretrained-era version is prompt learning: freeze the whole backbone and learn only a small pool of retrievable "prompt vectors" (L2P [S104], DualPrompt [S105]) — shrinking "capacity for the new task" to almost no parameters at all.
• Fast learning + slow consolidation = Progress & Compress: an active column learns the new task quickly, then is distilled (one network is trained to imitate another's outputs, compressing its knowledge across) into an EWC-protected knowledge column [S57].
• A literal machine version of sleep. Tadros et al. (2022) inserted a "sleep phase" (noisy input + local Hebbian rules — Hebb's principle: "neurons that fire together, wire together") into artificial neural networks; spontaneous replay recovered previously forgotten tasks and made representations sparser and less overlapping [S82]. Hinton et al.'s wake-sleep algorithm (1995) prefigured the idea [S81]. [established (each result); the general engineering value of sleep mechanisms: speculative]

4.3 Keeping knowledge outside the weights: external memory and retrieval

A different move: if writing new knowledge into weights is this dangerous, don't write it there.

• Differentiable external memory ("differentiable" meaning the whole read/write process can itself be trained by gradients): the DNC showed a network can learn to read and write an external store [S63].
• kNN-LM: retrieving nearest neighbors at inference substantially improves language modeling — parametric knowledge plus non-parametric memory is a naturally effective hybrid [S65].
• RAG: retrieval-augmented generation is now the default industrial answer for injecting new or private knowledge into LLMs — updating knowledge = updating the document store; zero forgetting, zero training [S64].

The limits are equally clear: retrieval is excellent for declarative knowledge (facts, documents), but procedural capability (skills, reasoning patterns, taste) does not live at a retrievable granularity; unbounded store growth needs management; and "retrieved but poorly used" integration failures point back to needing some consolidation mechanism that compresses frequently used memories into the weights. [established (effectiveness); the limit analysis is synthesis]

4.4 In-context learning: every forward pass is a small continual-learning episode

GPT-3 established ICL: with no weight updates, examples in the context suffice for the model to "learn" a new task [S69]. Mechanistic work shows this is more than lookup: a transformer's forward pass can implicitly implement gradient descent [S70], and in controlled linear settings ICL implicitly executes standard learning algorithms [S71].

This gives lifelong learning an important lens: the context window is a zero-forgetting, write-instantly, evaporates-after-the-session fast memory — functionally, strikingly like the hippocampus's role. Its three ceilings: capacity (window length), persistence (gone when the session ends), and cost (quadratic attention). Test-time training aims to make that evaporating memory durable: TTT (2020) updates the model at test time with a self-supervised loss (self-supervision: exercises constructed from the data itself, needing no human labels) [S66]; TTT layers (2024) make a recurrent network's internal state itself a small model trained by gradient steps during inference [S67]; Titans adds a surprise-gated neural long-term memory with adaptive forgetting, scaling beyond contexts of 2 million tokens (a token is the basic unit of text a model processes, roughly a word or part of one) [S68]. [established (each result); "this is the embryo of a lifelong-learning architecture": speculative]

4.5 Continual pretraining and fine-tuning of LLMs: the closest thing to a practical battlefield

• Domain-adaptive continual pretraining (DAPT/TAPT) shows consistent gains from "just keep pretraining" [S72]; Ke et al. give a soft-masking method for forgetting-resistant continual pretraining [S73]; Gupta et al. (2023) systematically studied the key ingredient of learning-rate re-warming [S119].
• Ibrahim et al. (2024), in a large-scale study, provide the most practical recipe to date: re-warm then re-decay the learning rate + mix in a small fraction (~5%) of old-distribution data, which approaches "retrain from scratch on the union" performance across hundreds of billions of new tokens, under both weak and strong (English→German) shifts [S74]. This is replay plus plasticity management winning at frontier scale. [established]
• On the fine-tuning side: instruction-tuned models sequentially learning many tasks keep their abilities with as little as ~1% replay (Scialom et al. 2022 [S75]); yet 1B–7B models show significant forgetting under continual fine-tuning, worsening with size in that range (Luo et al. 2023 [S76]) — the relationship with "scale/width mitigates forgetting" [S30][S106] is unresolved. [contested]
• Knowledge editing (hypernetwork-style MEND [S121]; locate-and-rewrite ROME/MEMIT [S77][S78]) can surgically rewrite factual associations, but its failure at scale is now precisely characterized: as sequential edits accumulate, the model first suffers gradual forgetting (declining editability, loss of earlier edits, sliding downstream performance), then a single "disabling edit" triggers abrupt catastrophic collapse (Gupta et al. 2024 [S120]). Surgical writes do not currently accumulate without bound. [established]
• For the full landscape, see the two surveys of continual learning for LLMs [S125][S126].

4.6 Agents that learn from experience

Voyager demonstrated lifelong learning that sidesteps weight updates: skills discovered in Minecraft are distilled into a code skill library, reused and composed to keep getting stronger [S79]; Reflexion writes failures into verbalized reflections stored in episodic memory [S80]. This "experience → text/code → retrieval" loop is currently the mainstream way LLM agents accumulate capability. Its ceiling is the same as §4.3: capability accumulates outside the weights while the base model never grows. Silver and Sutton's "Era of Experience" [S20] argues for crossing exactly that threshold — agents learning directly (including weight updates) from lifelong experience streams — which pushes every difficulty in this document (forgetting, plasticity, consolidation) back to center stage. [established (the system demos); the strategic judgment is a published position the authors share]

Part 5: How Brains Avoid Forgetting — and Which Mechanisms Transfer

Biological learning systems run for decades without catastrophic forgetting; they are the only existence proof that lifelong learning is possible. Mechanism inventory (overview: Kudithipudi et al. 2022 [S22]):

5.1 The synaptic level: memory written across timescales

• Spine stability. Dendritic spines are the tiny protrusions on which neurons receive most of their synaptic connections. A fraction of learning-induced spines persist for life, forming the structural substrate of long-term memory (Yang et al. 2009 [S83]); plasticity for different tasks is allocated to different dendritic branches, a structural isolation at subcellular scale (Cichon & Gan 2015 [S85]). [established]
• Cascade / multi-state synapses. The models of Fusi et al. (2005) and Benna & Fusi (2016) show that if each synapse contains a chain of fast-to-slow hidden variables, memory capacity and lifetime improve enormously, with power-law rather than exponential forgetting [S89][S90]. This is one of the most concrete lessons for deep learning: a scalar weight is too impoverished — every weight should carry multi-timescale internal state. (EWC's penalty, optimizer momentum, even L2 Init can be read as crude approximations.) [established (the models); the transfer judgment is inference]
• Metaplasticity. Synapses change how changeable they are (Abraham & Bear 1996 [S91]) — the learning rate is not a global hyperparameter but a regulated per-synapse state variable. [established (biology); machine analogues scattered across methods]

5.2 The systems level: consolidation is an active offline process

• The hippocampus replays waking activity sequences during sleep (Wilson & McNaughton 1994 [S86]) — direct evidence for CLS [S61][S62]; consolidation takes a time window during which it can be disrupted (McGaugh 2000 [S87]). [established]
• Sleep does constructive work (branch-specific formation of new spines after learning, Yang et al. 2014 [S84]) and possibly global normalization (the synaptic homeostasis hypothesis, SHY: global synaptic downscaling during sleep, restoring the next day's plasticity budget, Tononi & Cirelli 2014 [S88]). The tension between SHY and selective strengthening remains debated within neuroscience. [contested (the strong version of SHY)]
• Translation to AI: training should not be a single online pass; systems need periodic "offline housekeeping" phases — replay, distillation, compression, regularizing resets. Tadros et al.'s sleep-like replay [S82], the Compress phase of Progress & Compress [S57], even the replay mixture in LLM continual pretraining [S74] are fragmentary implementations of this principle. [inference]

5.3 The coding level: sparsity, separation, gating

• Sparse coding is a basic sensory-cortex strategy (Olshausen & Field 1996 [S92]); the dentate gyrus's pattern separation further de-overlaps similar experiences. Lower representational overlap → less interference — dovetailing exactly with the ML-side finding that forgetting is driven by representational overlap [S3][S29].
• Direct transfer successes: XdG's sparse context gating [S60] and OWM's context-dependent processing module [S45] were explicitly inspired this way and delivered results at hundreds-of-tasks scale. [established]
• The cost is real too: sparsity/separation trades off against generalization and transfer — fully non-overlapping representations cannot share knowledge. The brain's apparent answer is a division of labor: extremely sparse hippocampus, highly overlapping cortex, fast and slow. [established (the trade-off exists); the "division of labor" phrasing is a summary]

5.4 The neuromodulatory level: when and how fast to learn is itself a controlled signal

Neuromodulators (norepinephrine, dopamine) gate plasticity by surprise, novelty, and reward (Aston-Jones & Cohen 2005 [S99]). Machine counterparts are emerging: Titans writes to memory in proportion to "surprise" [S68]; importance-weighting schemes [S41][S42][S43] are static approximations. Promoting the learning rate from hyperparameter to a per-parameter, per-moment control signal estimated by the system itself is a direction strongly suggested by biology and not yet systematized in ML. [inference/speculative]

5.5 Neurogenesis: an honest cautionary tale

Adult hippocampal neurogenesis was long a poster child for "inject fresh plastic units" (Eriksson et al. 1998 [S93]). But in 2018, Sorrells et al. (Nature) could not detect new neurons in adult samples [S94], while Boldrini et al. (Cell Stem Cell) reported neurogenesis persisting to age 79 [S95] — a methodological dispute (postmortem delay, marker degradation, stereology) still not fully settled. [contested]

Two lessons: (1) whenever "the brain does X" is used to justify an AI design, the current state of the biological evidence must be checked; (2) intriguingly, whatever the biological facts, "continually injecting fresh, uncommitted units" has been independently validated on the machine side — precisely what continual backpropagation [S23] and dormant-neuron resets [S27] do. Engineering need not wait for biology to finish arguing.

5.6 The brain is not a flawless control group either

Humans forget too, and also show order effects — in the opposite direction: Flesch et al. (2018) found humans learn better under blocked training (finish A, then B) than interleaved, the reverse of standard neural networks [S96]. The brain is not "a neural network that resists forgetting" but a different kind of system with strong priors, gating, and consolidation. The right use of neuroscience for AI is extracting computational principles (multiple timescales, offline consolidation, sparse gating, neuromodulatory control), not copying parts (Hassabis et al. 2017 [S97]; Richards et al. 2019 [S98]). [established (the experiment); the methodological stance is a published position]

Part 6: Open Fundamental Problems (Ranked) and the Most Promising Directions

Eight problems, ranked by how much they bottleneck lifelong-learning AI (the ranking itself is the authors' judgment [speculative]; each item's evidence base is tagged separately). Each entry: why it is fundamental, current best evidence, and one concrete falsifiable research step.

P1. What is the objective function of lifelong learning? — the theory gap

• Why fundamental: the field lacks consensus even on what a continual learner should optimize. Approximate the joint multitask solution? Minimize long-run regret (the accumulated gap between each decision and what would have been optimal in hindsight) under compute/memory constraints? Maximize knowledge compounding (the acceleration of future learning)? Or, when the objective itself drifts, permit and reward correct forgetting (the interference-vs-revision distinction of §1.4; a published argument for this appears in the recommendation setting [S117])? Different answers yield entirely different methods and evaluations. The current stand-in — "average accuracy doesn't drop" — can be cracked by a method that does no continual learning (GDumb [S17]) and does not survive compute-budget matching [S124].
• Best evidence: Knoblauch's NP-hardness + perfect-memory theorem is one of the few theoretical foundation stones [S18]; the FWT/BWT metric framework [S15]; the evaluation critiques [S16][S17][S124].
• Falsifiable step: build a benchmark whose primary metric is learning acceleration — over a long stream (≥500 tasks), measure n(k), the samples needed for task k to reach threshold performance, and test the hypothesis that no existing method makes n(k) decrease persistently in k. If some method does, it is the first demonstrated case of knowledge compounding; if all fail, the field has made zero progress on its most important axis — either outcome is decisive.

P2. Necessary and sufficient conditions for plasticity loss — is randomness required?

• Why fundamental: if Dohare et al.'s strong claim ("gradient methods are not enough; a non-gradient random component is required" [S23]) holds, every purely gradient-based continual-learning route has a ceiling, and optimizers/architectures need redesign.
• Best evidence: the Nature 2024 experiments [S23]; mechanism studies (dead units [S27], curvature/rank [S28]); the deterministic counterexample candidate L2 Init [S31].
• Falsifiable step: on the original Continual ImageNet and drifting-regression protocols, systematically compare deterministic methods (L2 Init, per-parameter adaptive regularization, second-order methods) against random-reset methods out to ≥5,000 tasks. The hypothesis "a purely deterministic optimizer can maintain plasticity indefinitely" is directly decidable — and important either way.

P3. The capacity limit of "learning nearby" — how much knowledge fits in one basin?

• Why fundamental: Part 3 shows today's most practical route (LoRA, merging, task arithmetic, orthogonal projection, continual merging) rests entirely on "new knowledge fits near the old solution." Nobody knows after how many tasks, or how much distribution shift, the assumption fails — or whether the failure is predictable.
• Best evidence: linear connectivity [S34][S40]; merging efficacy [S48][S49][S115][S116]; the "learns less, forgets less" low-rank trade-off [S113]; the shrinking-subspace arithmetic [S46][S114]; recent continual-merging methods and their self-reported limits [S117][S118]; regularization failure in class-incremental settings [S13]; the "loss-connected ≠ mechanistically same" caveat [S112].
• Falsifiable step: from a fixed pretrained model, apply sequential low-rank/orthogonal updates and directly measure (a) remaining usable subspace dimension, (b) the gap to a full-retraining solution, and (c) when linear connectivity breaks — as functions of task count and shift strength. This yields the first "basin capacity curve." The hypothesis "the exhaustion point is predictable in advance from spectral quantities (the eigenvalue distributions of the Fisher/NTK matrices)" is directly testable; both parameter-space and function-space distances should be reported (see §3.4).

P4. The division of memory and the consolidation policy — what goes in weights, what in external stores, and when to transfer?

• Why fundamental: theory (the theorem in P1 [S18]) and biology (CLS [S61][S62]) both say memory must be hierarchical, but the consolidation policy — which external memories get distilled into weights, when, and how — is currently pure hand-engineering. It is the missing piece for upgrading RAG-style systems into genuinely growing systems.
• Best evidence: replay's robustness [S13]; RAG/kNN-LM effectiveness [S64][S65]; the sleep-consolidation proof of concept [S82]; the LLM continual-pretraining recipe [S74].
• Falsifiable step: at fixed total compute, on a long-horizon streaming benchmark mixing QA and skills, compare pure RAG vs. pure continual fine-tuning vs. RAG + periodic "sleep" consolidation (distilling frequently retrieved items into the weights). The hypothesis "the hybrid beats both pure options at equal budget" is decidable; if pure RAG never loses, that would show weight consolidation may be unnecessary for declarative knowledge — an equally major conclusion.

P5. The science of continual pretraining at scale — a scaling law for forgetting

• Why fundamental: for LLMs, the realistic form of lifelong learning is continual pretraining. We have recipes (re-warming [S119] + replay [S74]) but no laws: how does forgetting scale with model size, replay ratio, shift strength, token budget? [S30][S106] say scale/width mitigate forgetting; [S76] observes the opposite trend within 1B–7B; the contradiction is unresolved. The editing route's scale ceiling is also now characterized (two-phase gradual-then-catastrophic degradation [S120]).
• Best evidence: [S30][S72][S73][S74][S75][S76][S106][S119][S120].
• Falsifiable step: across three orders of magnitude of model scale with controlled shift strength, fit a "continual scaling law": forgetting = f(scale, replay ratio, shift). The hypothesis "there exists a scale threshold beyond which forgetting → 0 at 5% replay" is directly testable and would adjudicate [S30] vs. [S76].

P6. Automatic distillation of experience into competence — how do agents turn episodes into skills?

• Why fundamental: today's agents "learn" only into text/code memory (Voyager [S79], Reflexion [S80]); the base weights never grow. The Era-of-Experience program [S20] requires crossing that gap. What is missing is a mechanism that automatically extracts procedural competence from episodic memory and writes it safely into weights — i.e., a complete machine implementation of CLS.
• Best evidence: capability accumulation in skill-library agents [S79]; consolidation-by-distillation prototypes [S57]; the scale ceiling of knowledge editing [S77][S78].
• Falsifiable step: in a long-horizon environment (hundreds of tasks in Minecraft or a robotics curriculum), compare a frozen-base + skill-library agent against one that additionally distills skills into its weights periodically. Hypotheses: the distilled agent retains most competence after the skill library is removed, and its learning speed on new skills increases with accumulation (connecting to P1's metric). If distillation always destroys more than it deposits, in-weight consolidation is infeasible under current architectures — also a decisive result.

P7. The interference geometry of representations — can we predict in advance what an update will destroy?

• Why fundamental: the microscopic anatomy of forgetting (which layers, which directions, how task similarity matters) is partially mapped [S29][S30], but there is no general prospective theory: given the current model and new data, predict which old capabilities will be damaged. With one, continual learning could shift from after-the-fact repair to before-the-fact avoidance.
• Best evidence: the theoretical starting point already exists — the NTK overlap matrix quantifies inter-task interference in the (approximately) linear regime [S107], with companion generalization guarantees for OGD [S108] and exact forgetting characterizations in linear regression [S109]; empirical anatomy at the representation level [S3][S29][S30][S39]. What is missing is pushing these from linear/controlled settings into the genuine feature-learning regime and LLM scale.
• Falsifiable step: test the hypothesis "overlap between new and old tasks in NTK-feature/representation space (the quantity of [S107] or a variant) predicts post-fine-tuning forgetting before training, with correlation significantly exceeding a parameter-distance baseline" — once on controlled task families, once on real LLM fine-tuning. If it holds, an interference-aware data-scheduling algorithm falls out directly.

P8. Evaluation itself — an honest benchmark for open-ended lifelong learning

• Why fundamental: every step in P1–P7 needs measurement instruments that cannot be gamed. The distortions of existing benchmarks are documented [S16][S17], and many gains vanish under budget matching [S124]; public benchmarks that are task-free [S19], single-pass online [S122], compute-accounted [S55][S124], and include learning-acceleration metrics remain scarce.
• Best evidence: [S16][S17][S19][S55][S122][S124]; Dohare's Continual ImageNet and drifting regression provide templates on the plasticity side [S23].
• Falsifiable step: build a public long-stream benchmark satisfying those constraints, reporting ACC/BWT/FWT/learning-acceleration plus compute and memory bills, then test the hypothesis that method rankings on standard benchmarks differ significantly from rankings on this one. A large rank flip (a recurrence of the GDumb phenomenon) would prove the field's progress metrics have been systematically distorted.

The most promising directions (authors' ranked synthesis; all [speculative]-grade judgments, reasons and evidence as noted)

1. Complementary learning systems for LLMs: retrieval-based episodic memory (instant writes, zero forgetting) + periodic offline consolidation (replay/distillation into weights). Rationale: both components are individually validated ([S64][S74][S82]); only the closed loop is missing; it addresses P4 and P6 simultaneously; and it is compatible with existing LLM infrastructure, so industry can push it immediately.
2. Plasticity-preserving optimization and architectures (the continual-backpropagation family, per-parameter metaplasticity, multi-timescale weight states [S23][S31][S90]): no long-running system escapes P2; this is foundation-grade work.
3. Recipe science and scaling laws for continual pretraining (P5): closest to practice, cheapest to generate data points for, and every finding transfers directly to production systems [S74].
4. Architectures with learning built into inference (TTT layers, Titans-style neural memory [S67][S68]): success here would dissolve the train/deploy two-phase split — the foundational defect named in Conclusion 1. The risk: no lifelong-scale evidence yet.
5. Geometry-guided modular updates and merging (LoRA + task arithmetic + aligned merging [S47][S48][S36]): certain short-to-medium-term payoff, but bounded by P3's capacity limit — more likely a transitional technology than the endgame.

One sentence to close: catastrophic forgetting and loss of plasticity are not bugs to patch but inevitable consequences of one design decision — storing all knowledge in densely shared parameters with a single timescale. The way out is not cleverer patches but the layered architecture brains already use: parameters with multiple timescales, fast and slow memory systems with a division of labor, active offline consolidation, and an objective function that honestly measures the compounding of knowledge.

Method and Limitations

• This is primarily a literature synthesis and argument; the only experiment of our own is the toy demo at the end of Part 3, explicitly labeled illustrative and not used as evidence. The original coverage ends in 2025; the 2026-09 update added a batch of individually verified public references (SOURCES.md section F), including continual-merging / proximally anchored low-rank adapters, NTK interference theory, budget-matched evaluation, and the characterization of model editing's failure at scale. Continual learning is a fast-moving field, and Conclusions 3, 9, and 12 are the most likely to need revision as evidence accumulates.
• Citation policy: arXiv identifiers and DOIs are authoritative for every reference; conference/journal names are stated only where directly verified (official proceedings or conference pages), and omitted otherwise.
• All rankings (Part 6) are the authors' synthesis over the evidence and are uniformly tagged [speculative]; treat them as debatable positions, not facts.
• All references are publicly available; entries verified online during writing are noted at the end of SOURCES.md. Where any citation is found to be in error, the original literature governs.
• Chinese primary version: STUDY.zh.md; one-page summary: SUMMARY.zh.md.