Introduction: Lifelong Learning and the Problem of Forgetting

Lifelong learning is the ability of an AI or robot to continually acquire new skills and knowledge over time without losing the skills it has already learned. A notorious challenge in this paradigm is catastrophic forgetting – when learning new tasks abruptly erases or distorts previously learned knowledge . In robotics, this can be dire: imagine a service robot that learns a new kitchen task only to forget how to open the fridge or where items are located, causing it to fail at multi-step routines. Researchers have proposed selective forgetting as a possible solution – intentionally removing or suppressing less important older knowledge to make room for new learning. However, as we will explore, this strategy may be at odds with how human memory works. Humans are capable of lifelong learning without such severe trade-offs, thanks in part to robust procedural memory and subconscious consolidation processes . This post critiques selective forgetting in robotic learning by examining the LIBERO benchmark as a case study and contrasting it with insights from cognitive science and neuroscience.

LIBERO: A Case Study in Lifelong Robot Learning

Figure 1: The LIBERO benchmark (Lifelong Robot Learning) provides a suite of 130 robot manipulation tasks organized into four task sets (Spatial, Object, Goal, and a 100-task mixed set). Each set isolates different types of knowledge transfer (spatial layouts, object variations, goal variations, or entangled mixtures). LIBERO is designed to evaluate how algorithms handle declarative knowledge (facts about objects/goals) vs procedural knowledge (how to perform actions) in sequential learning.

The LIBERO benchmark was introduced to spur research in lifelong robot learning, especially continuous decision-making tasks that mix declarative and procedural knowledge. It poses challenges like distribution shifts between tasks, different task orderings, and requires knowledge transfer rather than training each task in isolation. Crucially, LIBERO emphasizes the need to prevent forgetting of early tasks while learning new ones, echoing the stability–plasticity dilemma in AI. To tackle this, LIBERO provides baseline implementations of three representative lifelong learning methods : • Experience Replay (ER): a memory-based method that stores a subset of past experiences and replays them during training on new tasks . By interleaving old task data with new, ER aims to refresh the network’s memory of earlier skills and avoid forgetting. This is analogous to giving the robot “flashcards” from prior tasks while it learns a new task. ER is effective but incurs extra memory overhead, since some data from each task must be retained . • Elastic Weight Consolidation (EWC): a regularization-based method that protects important knowledge by slowing down learning on parameters crucial to old tasks . EWC computes an importance score (using the Fisher information matrix) for each neural weight based on how vital it was to previous tasks . When a new task is learned, changes to high-importance weights are penalized, thereby preserving previously learned skills. In essence, EWC implements a selective memory retention – it doesn’t forget so much as it “remembers” certain parameters strongly, reducing their plasticity. • PackNet: a dynamic architecture method that performs iterative pruning of the neural network to free capacity for new tasks . When learning a sequence of tasks, PackNet will selectively forget by pruning away the “less important” neurons (weights) after each task, based on which parts of the network were least used or needed . The pruned neurons are then reinitialized and allocated to learning the next task . By assigning disjoint subsets of the network to different tasks (through binary masks), PackNet effectively creates isolated skill modules, achieving minimal interference (and thus minimal forgetting) between tasks . However, this comes at a cost: because each task only uses a fraction of the network, the model’s capacity for any single task is limited, and knowledge cannot be easily shared across tasks . PackNet is an explicit implementation of selective forgetting – it decides what to “forget” (prune) and what to keep, to balance old and new knowledge.

These methods correspond to prevalent strategies in lifelong learning: rehearsal (replay), regularization, and modular expansion/pruning. In LIBERO’s experiments, one might expect that such sophisticated algorithms outperform naive training. Interestingly, the LIBERO team found a surprising result: plain sequential fine-tuning (simply training tasks one after another without any special forgetting-prevention) outperformed these lifelong learning methods in terms of learning new tasks (forward transfer) . In other words, methods like EWC and PackNet succeeded in preserving prior knowledge (preventing forgetting) but at the expense of slower adaptation to new tasks . This led to worse performance on the newest tasks compared to a model that freely adapted (and unfortunately, likely forgot some old tasks) . This empirical finding highlights a core tension: focusing on not forgetting can constrain the ability to acquire fresh skills. The selective forgetting approach in PackNet, for instance, minimized forgetting but also walled off each task’s knowledge, reducing beneficial transfer and eventually hitting a capacity wall . In a 100-task sequence, such rigid compartmentalization is not how humans would handle learning. To understand why, we turn to how humans leverage memory in lifelong learning.

Human Procedural Memory and Subconscious Learning

Humans are remarkably adept at learning new skills throughout life while retaining old ones. A key to this ability lies in our procedural memory – the memory system for skills and habits. Procedural memory is a form of implicit memory, meaning we use it without conscious awareness . For example, once you’ve learned to ride a bicycle or type on a keyboard, you can perform these activities without actively recalling when or how you learned them. The knowledge has been encoded subconsciously. This implicit nature of procedural learning is powerful: it lets us layer new skills on top of old ones seamlessly. As one text puts it, “Procedural memory… aids the performance of tasks without conscious awareness of previous experiences.” . In other words, our brains integrate the experience of practice so deeply that executing the skill becomes automatic (often called “muscle memory”).

Importantly, procedural memories tend to be robust and long-lasting. Even individuals with severe memory impairments can retain procedural skills. Classic studies on amnesic patients (like the famous patient H.M.) and Alzheimer’s patients show that they can learn and retain motor skills such as mirror-tracing, even when they cannot remember having learned the task or recall factual information about the training sessions . In one study, amnesic patients demonstrated normal acquisition and long-term retention of a new visuomotor skill, indicating that the skill-memory was intact despite their inability to form new declarative memories . This implies that procedural memory is housed in different neural circuits (e.g. basal ganglia and cerebellum for motor skills) separate from the brain’s declarative memory (hippocampus and cortex for facts and events). The architecture of the brain’s memory systems inherently shields skills from the kind of catastrophic forgetting that plagues single neural networks. A person might forget the name of a classmate (declarative memory) but not forget how to walk, tie shoes, or play the piano even after years of not practicing – procedural knowledge endures, in part because it becomes encoded in subconscious routines.

Another human advantage is the role of sleep and offline consolidation in learning. The brain does not necessarily learn tasks in one go; it reprocesses experiences during rest. Neuroscience research suggests that the brain replays or reorganizes memories during sleep, which helps integrate new learning without overwriting old memories. This is the brain’s way of solving the stability–plasticity dilemma. During sleep, recent experiences (including new skills) are gradually consolidated into long-term memory stores, a process thought to strengthen important neural connections while letting unimportant transient memories fade. Notably, this replay is selective – the brain doesn’t literally review all past experiences every night. Instead, it seems to reinforce salient experiences and blend them into existing knowledge networks . Through this subconscious process, humans achieve something that AI algorithms struggle with: we learn interleaved with periods of consolidation, rather than continuous overwriting. As a result, the human brain is capable of lifelong learning of new tasks without interfering with its ability to perform previously learned tasks . We gain new skills (like learning a musical instrument, then a new language, then a new sport) and usually accumulate abilities instead of replacing them. Forgetting does happen in humans, of course, but it is usually gradual and tied to lack of use or relevance, not an intentional mechanism to free capacity for a new skill.

It’s also worth noting that humans leverage a mixture of explicit and implicit learning. We consciously recall facts or instructions (explicit memory) when needed, but much of our skill acquisition happens implicitly by doing and practicing. This creates a layered knowledge base: a conscious understanding (which can fade or be limited) and a deeper habit memory. For example, when you learn to drive a car, you might memorize rules and tips explicitly at first, but eventually you operate the vehicle with very little conscious thought – your feet and hands “just know” what to do. This subconscious proficiency means new learning (say you later learn to drive a truck) can often build on the old skill without erasing it. You don’t have to forget how to drive a car in order to learn to drive a truck; you adapt your existing skillset and form a specialized extension of it. The human brain’s modular yet interconnected memory systems allow reuse of foundational abilities in new contexts, rather than starting from scratch each time.

Where Selective Forgetting Falls Short

When we compare the above characteristics of human learning to selective forgetting strategies in robotics like those in LIBERO, we start to see why the latter can be insufficient or oversimplified. Selective forgetting algorithms treat memory as a zero-sum resource: to learn something new, delete or suppress something old (ideally the “less important” pieces of knowledge). This approach is a direct response to the limitations of a single artificial neural network that must encode everything. But the human brain doesn’t usually operate in such a strict zero-sum fashion. Instead, it has mechanisms to expand and reorganize memory. If one neural circuit becomes saturated with skills, the brain can recruit additional circuits or refine existing ones (think of how learning a second language might involve engaging new regions, or how practicing a skill can make its neural representation more efficient). In AI terms, the brain is more like a dynamic, plastic architecture that grows and reconfigures, rather than a fixed-size network that must prune old connections to learn new ones . Indeed, some researchers have argued that the mammalian brain avoids catastrophic forgetting by protecting old knowledge in certain circuits while still allowing new learning in others . This is analogous to having a built-in form of EWC or PackNet, but far more flexible: the brain can isolate critical memories (stability) and allocate new resources for novel tasks (plasticity) as needed, largely without explicit “forgetting protocols.”

Selective forgetting methods like PackNet or EWC are engineered simplifications of these ideas – they try to mimic the brain’s compartmentalization or synaptic consolidation. But in doing so, they often miss the fluidity of human memory. For instance, PackNet prunes weights to prevent interference, yet this rigidly partitions knowledge and prevents any positive transfer between tasks . Humans, on the other hand, often transfer prior skills to new ones (we call this analogical reasoning or generalization). A pianist learning the guitar doesn’t prune away their knowledge of rhythm; they carry it over, while learning new finger motions. In AI, if we completely isolate the “piano network” from the “guitar network” to avoid forgetting, the guitar network might not benefit from the rhythm knowledge in the piano network. LIBERO’s finding that sequential fine-tuning (which allows some forgetting but also more sharing of representations) had better forward transfer than strict forgetting-prevention methods reflects this conundrum . By preventing forgetting too rigidly, current methods also prevented the subconscious blending of knowledge that could accelerate new learning.

Another oversimplification is the notion of actively choosing what to forget. In human cognition, forgetting is usually not a deliberate, targeted act (except perhaps when we consciously ignore irrelevant info). It’s an emergent property – a byproduct of our brain updating and reorganizing memories. We do forget, but often the brain forgets specifics (e.g. exact details that are unused) while retaining the gist or skill; this is believed to help in abstracting and generalizing. Robotic selective forgetting, in contrast, might drop entire chunks of learned parameters to free capacity. This could throw away subtle correlations or sub-skills that might have been useful. The granularity of human forgetting is different: we lose resolution (details) before core abilities. Current AI methods struggle with this kind of nuanced forgetting. They either remember a task nearly perfectly or not at all, with little in-between. For example, a neural net either retains the ability to solve Task A or it is almost completely overwritten by Task B, unless mechanisms like replay or regularization intervene . Humans rarely exhibit such all-or-nothing forgetting; an old skill might get rusty, but not vanish instantaneously after learning something new.

From a theoretical standpoint, selective forgetting approaches address symptoms (catastrophic forgetting) but not the root cause: the monolithic nature of most AI knowledge representations. If we truly mimic human learning, we might need AI systems with multiple memory systems (analogous to our separate procedural and episodic memory stores, or short-term vs long-term memory) and with periodic consolidation phases (like “sleep” for neural networks). Some research is heading in this direction – for instance, using episodic memory modules that store key past experiences (like a hippocampus) and gradually distill them into the main model (like the cortex), or having “sleep” phases where the AI replays memories to strengthen them . These approaches acknowledge that simply chopping out memories (as in selective forgetting) is too crude. The human brain’s success comes from layered learning (conscious + subconscious, short-term + long-term) and structural separation of knowledge, rather than a manual delete button for memory. Selective forgetting in robots, as implemented today, doesn’t capture these richer dynamics.

Philosophical Reflections: Bridging Robotic and Human Learning

The gap between robotic learning algorithms and human learning is not just technical – it’s deeply philosophical. It raises questions about the nature of memory, knowledge, and even identity in machines versus humans. A human’s memories and skills define who they are; forgetting can be traumatic or freeing, but it’s often involuntary. In robots and AI, we are designing the memory process from the ground up. When we implement selective forgetting, we are essentially playing the role of nature and nurture combined, deciding what a machine should remember or discard for optimal performance. This leads to a thought-provoking point: Should machines forget in the same way humans do, or should they not forget at all? If an AI could retain everything perfectly, it wouldn’t need to forget – but then, would it drown in an ocean of data? Humans need to forget certain things (e.g. trivial details) to focus on what matters; perhaps intelligent machines also need a form of forgetting to remain efficient. The philosophical twist is that in AI we can tailor forgetting, whereas in humans it evolved as a necessity. Are we oversimplifying the rich tapestry of human memory by reducing it to a tunable parameter?

The LIBERO benchmark’s focus on lifelong learning algorithms brings us to ask: what kind of mind are we trying to build in our robots? The selective forgetting approach views the mind as a finite container that must be managed; the human brain suggests a view of mind as an ever-expanding, self-organizing web. The fact that current robots struggle with lifelong learning indicates that we might need a paradigm shift. Perhaps the goal shouldn’t be to selectively forget, but to selectively remember in a more human-like way – prioritizing important skills but integrating new ones through consolidation. Instead of hard resets or pruning, a robot might benefit from an analog of sleep, dream, or play: processes that appear in human cognition to reinforce and interconnect memories. Indeed, cognitive scientists have noted that play and exploration in humans lead to deeper learning; if a robot continuously learns in a utilitarian, task-by-task fashion (with bits of memory excised here and there), it misses out on the holistic development that humans undergo.

There is also an ethical dimension. If a lifelong learning robot is meant to interact with humans (as a personal assistant, for example), its ability to remember shared experiences is part of building a relationship. Would we want such a robot to “forget” earlier interactions to optimize for new tasks? That might make it a better task solver, but a poorer companion. Humans value continuity of memory – it’s tied to trust and personal connection. A robot that selectively forgets could behave inconsistently or lose the thread of its own “life story.” Philosophically, this touches on personal identity: continuous memory creates a narrative identity for humans; should we allow a robot’s narrative to be disrupted by algorithmic forgetting? Or should we instead develop memory systems that mirror the human balance of retention and loss, so that a robot’s learning journey is more organic and understandable?

In conclusion, the critique of selective forgetting in LIBERO and similar systems is not that avoiding forgetting is unimportant – clearly it’s crucial for robust AI – but that the method of doing so needs to evolve. The human brain has spent millions of years refining a solution to lifelong learning: procedural memories that persist, episodic memories that consolidate or fade gradually, and subconscious processing that binds it all together. Robotics and AI are just beginning to scratch the surface of these principles. By studying human memory – its resilience, its subconscious nature, and even its quirks like dreaming – we can inspire new algorithms that move beyond blunt selective forgetting. Bridging this gap is both a scientific and philosophical quest: it forces us to define what it means to learn and remember. Ultimately, closing the gap will not only make robots better learners but also bring them a step closer to the rich cognitive tapestry that biological intelligence enjoys. The journey to lifelong learning in robots may very well require us to embrace the very things that make human learning so enigmatic – the ability to remember, the necessity to forget, and the wisdom to know the difference.

Sources: 1. Bo Liu et al., “LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning,” NeurIPS Datasets and Benchmarks Track, 2023. 2. LIBERO Documentation – Lifelong Learning Algorithms (ER, EWC, PackNet). 3. Chongkai Gao et al., “PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning,” CVPR, 2018 . 4. James Kirkpatrick et al., “Overcoming catastrophic forgetting in neural networks,” PNAS, 114(13):3521–3526, 2017 . 5. IEEE Spectrum, “Sleep Can Keep AI From Catastrophic Forgetting,” 23 Nov 2022. 6. Wikipedia – Procedural memory (implicit skill memory, robustness in amnesia). 7. Nature News, “Why don’t new memories overwrite old ones? Sleep science holds clues,” 03 Jan 2025 . (Research on mice suggesting mechanisms to avoid catastrophic forgetting).