Memory, Learning, Conditioning and Behaviour: From the Life Sciences to Artificial Intelligence

Introduction: Understanding How Experience Transforms Intelligent Systems

Memory, learning, conditioning and behaviour are four fundamental domains for understanding how an organism acquires knowledge, adapts its responses to its environment and modifies its actions in the light of experience. Long studied separately by philosophy, experimental psychology, biology, physiology and the behavioural sciences, these phenomena now form a set of closely interrelated issues at the heart of research into artificial intelligence (AI).

Bringing these fields together is not merely a matter of drawing an analogy between the human brain and the computer. Rather, it involves a series of shared questions: how is information retained? How does experience modify future responses? How does a system distinguish relevant situations from incidental events? How does it adjust its actions according to the consequences it observes? Finally, how does a particular acquisition become a lasting skill that can be brought to bear in new circumstances?

These questions are central to contemporary neuroscience, machine learning, autonomous robotics, adaptive systems and AI architectures capable of exploiting information gathered through multiple interactions.

Nevertheless, it is essential to distinguish between mechanisms that, despite their interconnections, are not interchangeable. Memory refers to a set of processes involved in encoding, retaining and retrieving information. Learning involves relatively lasting changes in knowledge, abilities or dispositions to act. Conditioning examines certain forms of learning in which relationships between events, actions and consequences play a decisive role. Behaviour, finally, refers to the observable manifestations of an organism’s activity within a given environment.

Artificial intelligence reproduces certain functional structures associated with these phenomena, but implements them through computational mechanisms that differ profoundly from biological processes. A model can learn statistical regularities without possessing autobiographical memory comparable to that of a human being. An agent can optimise its actions on the basis of a reward without experiencing pleasure or frustration. A robot can modify its strategy after several attempts without thereby developing subjective experience.

Understanding these distinctions helps avoid two opposing mistakes: treating AI as a mere reproduction of the brain or, conversely, disregarding the lessons accumulated by the life sciences concerning the mechanisms of adaptation.

The history of these disciplines therefore provides a particularly fruitful framework for understanding the principles of artificial intelligence, from early associative models to deep learning systems, including reinforcement mechanisms, computer memory and architectures capable of interacting with the physical world.

1. The Scientific History of Memory: From Philosophy to Neuroscience

1.1. Memory as a Philosophical Problem

Memory has been one of philosophy’s major concerns since antiquity. It raises questions about knowledge, personal identity, the continuity of the self and the very possibility of experience.

In Plato’s philosophy, the theory of recollection links knowledge to the recognition of knowledge that the soul is presumed capable of recovering. This conception belongs to a particular metaphysical framework; it is not an experimental theory of memory in the modern sense. Nevertheless, it illustrates a fundamental question: what is the relationship between what is learned, what is retained and what becomes accessible to consciousness?

Other philosophical traditions have regarded memory as a faculty that preserves traces of sensory experience. In the nineteenth century, Henri Bergson notably distinguished between habit-memory, associated with repetition and the acquisition of movements, and recollective memory, which brings back particular events from the past.

This distinction does not correspond exactly to contemporary classifications of memory, but it highlights a problem that remains relevant today: not all forms of retention serve the same purpose. Knowing how to ride a bicycle, recognising a face and remembering a personal event involve different abilities, even though they interact in everyday life.

Memory therefore cannot be reduced to a uniform repository of information. It consists of a set of processes whose properties depend on the nature of the content, the conditions under which it was acquired, the time that has elapsed and the operations required to retrieve it.

1.2. The Emergence of Experimental Psychology

During the nineteenth century, the study of memory gradually began to move away from philosophical speculation and become an object of experimental measurement.

Hermann Ebbinghaus played a major role in this transformation. In his work published in 1885, he investigated the memorisation of syllables lacking any obvious meaning, thereby seeking to minimise the influence of pre-existing knowledge. He measured the effects of repetition, time and relearning on retention.

His experiments helped establish the forgetting curve: in the absence of reactivation, the amount of information that remains accessible generally tends to decrease over time. This process is neither uniform nor independent of the material being remembered, however. Meaning, attention, prior knowledge, sleep, emotions and opportunities for retrieval can substantially alter the trajectory of a memory.

Ebbinghaus’s historical significance lies less in establishing a universal law than in introducing a method: isolating certain variables, organising learning conditions and comparing results quantitatively.

This approach anticipates part of the methodology now used to evaluate artificial intelligence models. A recognition system, language model or robotic agent can be subjected to controlled tasks in order to measure its ability to retain information, generalise a rule or retrieve knowledge after a delay.

The comparison is methodological rather than biological: a computer system’s performance curve does not demonstrate the existence of a mnemonic mechanism identical to that of the brain.

1.3. Forgetting Is Not Simply the Disappearance of a Memory

Contemporary research has considerably refined earlier conceptions of forgetting.

Information that has become difficult to recall has not necessarily been erased. A failure of retrieval may result from a lack of appropriate cues, competition between memories, a transformation of the representation or a change in the conditions under which the information was acquired.

The distinction between availability and accessibility is essential here. Knowledge may not be readily available for spontaneous recall yet become accessible again when the context, the wording of a question or the presence of a cue facilitates retrieval.

Research into proactive and retroactive interference has also shown that new acquisitions do not always accumulate independently. Older information may disrupt the learning of new material; conversely, recent acquisitions may make it more difficult to recall previously learned content.

This complexity also arises in computer systems. A knowledge base may retain old information that has become obsolete. A model may generate an answer based on previously learned regularities even when the current context calls for an update. An agent may also become less effective when a new task interferes with the parameters acquired for an earlier one.

Biological forgetting, limitations in retrieving information and changes to a model’s parameters must nevertheless be distinguished. These phenomena may produce comparable functional consequences without relying on the same mechanisms.

1.4. Working Memory and Long-Term Memory Systems

Research during the twentieth century gradually led to the distinction between several memory systems.

Working memory refers to the mechanisms that temporarily maintain accessible information and allow it to be manipulated while a task is being performed. It is involved, for example, in mental arithmetic, understanding a complex sentence and comparing several options.

Long-term memory encompasses more enduring forms of retention. It includes episodic memory, which concerns personally experienced events; semantic memory, which is associated with general knowledge; and procedural memory, which contributes to the acquisition of skills and habits.

These categories should not be understood as perfectly isolated compartments. Real-world tasks generally recruit several processes, and theoretical classifications have evolved as cognitive neuroscience has advanced.

The distinction between declarative and non-declarative memory is also important. Declarative memory concerns knowledge that a person can generally express or report, whereas non-declarative memory encompasses certain skills, habits and forms of learning that do not require explicit conscious recollection.

Studies of patients with memory impairments have helped establish, in particular, that the ability to acquire certain skills may remain intact even when the formation of new declarative memories is severely disrupted.

This dissociation has transformed our understanding of memory: rather than depending on a single faculty, it relies on partially distinct systems whose interactions enable organisms to adapt their behaviour to a wide variety of situations.

1.5. From Consolidation to Neural Mechanisms

The biology of memory underwent a major transformation with the development of neurophysiology, molecular biology and techniques for recording brain activity.

Contemporary research shows that learning can alter the efficiency of connections between neurons, the excitability of certain cells, the organisation of neural networks and the expression of genes involved in synaptic plasticity.

Consolidation refers to a set of processes that stabilise certain acquisitions after they have been formed. It encompasses cellular and molecular mechanisms, as well as processes that reorganise neural networks.

Long-term potentiation, studied notably in the hippocampus, provides an important experimental model of lasting changes in synaptic efficacy. It does not, however, account for all the mechanisms of memory: several forms of plasticity are involved, depending on the circuits, tasks and learning conditions.

Sleep also plays a role in consolidating certain forms of memory. Interactions between the hippocampus and the neocortex, brain oscillations and the reactivation of particular representations help explain how recent acquisitions can be stabilised or reorganised.

Memories are therefore not necessarily stored as files located within individual brain cells. They depend on distributed changes across biological networks whose functioning evolves through experience.

This organisation offers a useful conceptual perspective for artificial intelligence. Modern models also rely on distributed representations: information can be encoded in the relationships among numerous parameters rather than in a single unit directly corresponding to a concept.

The similarity is nevertheless limited. The numerical parameters of an artificial network are not biological synapses, and modifying them does not entail the same energetic, cellular, developmental or physiological constraints.

1.6. The Myth of the Molecular Transfer of Memories

Some historical research conducted during the 1950s and 1960s sought to determine whether acquired information could be transferred from one animal to another through biological extracts.

Experiments associated in particular with James V. McConnell’s work on planarians and Georges Ungar’s work on brain extracts gave rise to considerable controversy. The more ambitious interpretations, according to which an isolated molecule could transmit a specific behavioural memory, did not establish a general mechanism for transferring memories.

Subsequent research confirmed the importance of molecular mechanisms in synaptic plasticity and memory consolidation, without validating the hypothesis of a universal molecular code capable of directly transferring a complex memory.

The distinction is essential: molecules participate in the processes of memorisation, but this does not mean that a memory can be equated with a single chemical substance containing all the information associated with an experience.

Artificial intelligence offers an instructive contrast. A digital model can indeed be copied with its parameters, subject to applicable technical and legal constraints. This operation transfers a computational state that can be used by a compatible system. It is not the biological equivalent of transferring a personal memory, because it reproduces neither a subject’s lived history nor all the bodily and contextual conditions that give meaning to memories.

2. Learning: From Experience to Lasting Changes in Abilities

2.1. Defining Learning Without Confusing It with Performance

Learning refers to a process through which experience produces a relatively lasting change in knowledge, skills or dispositions to act.

This definition calls for a fundamental distinction between acquisition and performance.

Performance refers to what an individual manages to accomplish under given conditions. It depends on learning, but also on temporary factors such as fatigue, motivation, stress, distraction, task difficulty, physiological state and familiarity with the context.

A student may have acquired a concept yet fail to recall it during an examination. An experienced musician may give a poor performance because of fatigue. An artificial intelligence system may possess the capabilities required for a task but fail because of ambiguous wording or an unfamiliar input.

Conversely, a one-off success does not always demonstrate lasting learning. A model may produce the correct answer by chance, by exploiting a superficial cue or because the relevant information is explicitly present in the context.

Rigorous assessment of learning therefore requires repeated observations, delayed tests and tasks designed to distinguish immediate memorisation from generalisation.

2.2. Nature and Nurture: Moving Beyond a Simplistic Opposition

The history of the behavioural sciences has been shaped by debate over the respective contributions of biological factors and environmental influences.

Contemporary approaches generally avoid reducing development to either of these two poles. Genetic characteristics, nervous-system maturation, early experiences, social interactions and material conditions all contribute to the development of abilities.

Biological structures do not constitute an exhaustive programme that determines every future behaviour. Rather, they establish constraints, predispositions and developmental possibilities that are realised in particular environments.

Language acquisition illustrates this interaction. Human linguistic development depends on biological capacities characteristic of the species, but it also requires exposure to sufficiently rich and accessible linguistic interactions.

Similarly, the acquisition of social norms, motor habits and perceptual categories depends on the interaction between initial predispositions and repeated experience.

In animals, ethology has shown that certain behaviours have innate components while remaining sensitive to experience. Recognition of conspecifics, exploration, parental behaviour and feeding strategies may combine biological predispositions with specific forms of learning.

The distinction between innate and acquired characteristics therefore does not correspond to an absolute separation. It is an analytical problem whose resolution depends on the species, the abilities under investigation and the experimental conditions.

2.3. The Main Forms of Learning

Learning encompasses several mechanisms that can operate together in the same situation.

Associative learning involves establishing relationships between events, stimuli, actions or consequences. Classical and operant conditioning are two major forms of this process.

Non-associative learning includes habituation and sensitisation. Habituation refers to a decrease in the response to a repeated stimulus that has no significant consequences. Sensitisation, by contrast, involves an increase in reactivity following certain experiences, particularly intense or aversive ones.

Observational learning makes it possible to acquire information from other people’s behaviour or from the consequences of their actions. It plays an important role in social interactions and can help reduce the cost of individual trial and error.

Cognitive learning involves representations, categories, conceptual relationships and problem-solving strategies. It is not limited to the mechanical repetition of a response.

Statistical learning makes it possible to extract regularities from distributions of events. Human beings exploit certain perceptual and linguistic regularities, while machine-learning models use mathematical mechanisms to identify structures in data.

Finally, reinforcement learning examines how an agent adapts its actions according to the consequences and rewards received during interactions with an environment.

These categories are not watertight compartments. Someone learning to play a musical instrument may simultaneously acquire motor habits, memorise sequences, adjust movements through sensory feedback and construct an abstract representation of musical structure.

Similarly, a robot may combine reinforcement learning, visual perception, state estimation and the use of previously acquired knowledge.

2.4. Contiguity and the Association of Events

Temporal proximity between events plays an important role in several forms of associative learning.

When a stimulus regularly precedes a significant event, an organism may learn to use that stimulus as a predictive cue. However, mere simultaneity is not sufficient to explain every association: predictability, contingency, prior experience and the characteristics of the subject also play a role.

In classical conditioning, a stimulus that was initially neutral may acquire the ability to elicit a response when it reliably predicts a biologically significant event.

In human learning, associations can extend beyond elementary perceptual relationships. Language makes it possible to connect concepts separated in time or space, describe hypothetical events and acquire knowledge through stories or explanations.

Language models also exploit relationships between units of text, but they do so through statistical and computational procedures. Their training involves adjusting parameters to minimise a loss function associated with a given task.

It would nevertheless be inaccurate to equate the learning of linguistic co-occurrences with the formation of a human psychological association. The relationship between the two depends on how representations are constructed, used and evaluated.

2.5. Repetition: An Important but Insufficient Condition

Repetition contributes to the stabilisation of many acquisitions. It provides additional opportunities to retrieve information, correct errors and improve the execution of a skill.

Its effectiveness, however, depends on how practice is organised.

Passive repetition followed immediately by another identical repetition can create an impression of mastery that exceeds actual retrieval ability. By contrast, active recall, varied exercises and delayed testing often promote more durable retention.

Spaced learning is one of the best-established findings in memory research. Distributing practice over time generally improves long-term retention compared with concentrating too many repetitions within a single period, although the optimal interval varies according to the material and the learning objectives.

Sleep, task difficulty, prior knowledge and opportunities to reuse acquired information also influence the results.

In artificial intelligence systems, repeated exposure to data can improve parameter adjustment, but raw repetition is not always beneficial. Redundant data may reinforce certain biases, lead to overfitting or increase computational costs without producing a proportional improvement in capabilities.

Effectiveness therefore depends on data quality, diversity, the learning objective and the training strategy.

2.6. Breaks, Consolidation and Interference

The timing of learning is as important as the total amount of practice.

Breaks can help restore attentional resources, limit certain forms of fatigue and allow new acquisitions to consolidate. Their optimal duration varies according to task complexity, individual characteristics and the objectives being pursued.

Interference also plays a major role. Two closely related learning experiences may reinforce one another when they share relevant structures, or hinder one another when their demands compete.

Positive transfer occurs when an acquired skill facilitates a new task. Negative transfer occurs when an established habit interferes with the acquisition of a different procedure.

In artificial intelligence, these problems arise notably in continual learning. A model trained successively on several tasks may lose certain earlier capabilities as its parameters are adapted to new data. This phenomenon is known as catastrophic forgetting.

Several approaches seek to mitigate this problem: retaining representative examples, regularising parameters, using modular architectures, selectively adapting parameters and implementing external memory mechanisms.

None of these methods exactly reproduces biological consolidation. They nevertheless address a common functional question: how can new capabilities be acquired without excessively degrading those that have already been learned?

2.7. Feedback and Error Correction

Feedback is an important factor in learning.

When a subject discovers the consequences of their actions, they can compare the result with what was expected, identify an error and adjust their strategy.

The effectiveness of this correction depends on the accuracy, timing and relevance of the feedback, as well as on whether the outcome can be linked to a particular action.

Feedback that is too vague may provide no actionable information. Immediate feedback can facilitate certain motor skills, whereas a delay or a period of reflection may prove useful for other tasks.

In supervised machine learning, the data generally provide examples associated with targets or annotations. The algorithm adjusts its parameters according to the discrepancy between its predictions and the reference results.

Reinforcement learning follows a different logic: the agent receives rewards or value signals associated with its interactions and seeks an action policy that improves its cumulative return.

These two methods share a logic of correction, but differ in the nature of the information available. An explicit label is not equivalent to a delayed reward, and neither guarantees an understanding of the task.

2.8. Knowledge Transfer and Generalisation

Learning has practical value only if it can be applied under conditions appropriate to its purpose.

Generalisation refers to the ability to produce relevant responses beyond the examples directly encountered during training. It requires the system to exploit sufficiently robust regularities to remain effective when faced with new variations.

In human beings, generalisation depends notably on prior knowledge, conceptual abstraction, an understanding of relationships and the ability to recognise the relevant features of a situation.

In machine learning, generalisation is assessed using data distinct from those used to train the model. Performance on a test set provides an indication of the ability to handle unseen examples, provided that those data adequately represent the conditions of use.

A model may nevertheless pass a test without being robust to distribution shifts, adversarial data or contexts far removed from those encountered during training.

Transfer is therefore a central challenge in modern AI. Excellent performance on a specialised task is not enough: systems intended to interact with the real world must handle variations, recognise their limitations and, when necessary, request additional information.

3. Conditioning: How Relationships Between Events Shape Responses

3.1. The Origins of Classical Conditioning

Classical conditioning is associated with the work of Russian physiologist Ivan Pavlov, who studied digestive processes before identifying anticipatory responses triggered by signals associated with food.

The principle rests on a learned relationship between two events.

An unconditioned stimulus initially elicits a response without requiring the experimental association under investigation. A neutral stimulus, by contrast, does not spontaneously elicit that specific response under the conditions being considered.

When the neutral stimulus regularly precedes the unconditioned stimulus, it may acquire a predictive function. It then becomes a conditioned stimulus, capable of eliciting a conditioned response.

The classic example involves pairing a sound with the presentation of food. Following appropriate learning, the sound may elicit an anticipatory response before the food arrives.

This transformation shows that behaviour depends not only on the immediate properties of stimuli, but also on learned relationships between events and their predictive significance.

3.2. Prediction Matters as Much as Temporal Proximity

Contemporary research has enriched the historical model of conditioning.

Temporal proximity between two events can facilitate their association, but it is not a sufficient explanation. An organism also appears to be sensitive to the informational value of a stimulus: a signal that reliably predicts an event may acquire greater significance than one that merely occurs at the same time.

This distinction is fundamental to understanding adaptive behaviour. In a complex environment, numerous events coincide without having a causal relationship. Learning must make it possible to distinguish relevant regularities from accidental associations.

Modern predictive algorithms face a comparable difficulty. They can exploit statistical correlations without correctly identifying the mechanisms that generate the data.

A model may learn, for instance, that a visual detail is frequently associated with a category, only to fail when that detail disappears or changes context.

The quality of predictive learning therefore depends on the robustness of the relationships being exploited and on the system’s ability to withstand variations that do not fundamentally alter the structure of the problem.

3.3. Extinction, Recovery and Generalisation

When a conditioned stimulus is no longer followed by the event it previously predicted, the conditioned response generally tends to diminish. This process is known as extinction.

Extinction does not necessarily mean that the original association has been erased. A response may reappear after a delay, in a different context or following another relevant experience.

Spontaneous recovery refers to the reappearance of a conditioned response after a period of time following extinction. Contextual renewal and reinstatement are other phenomena demonstrating that the history of learning continues to influence the response.

Stimulus generalisation occurs when a learned response to a particular signal also appears in response to similar stimuli. Discrimination refers to learning the relevant differences between several signals.

These mechanisms are important in the study of emotions, fear responses, habits and certain psychological difficulties.

They also have implications for artificial systems. An agent trained in a limited environment may generalise a rule to similar situations but fail when the context changes substantially. It may also retain a policy that has become unsuitable after the conditions of the problem have changed.

In both cases, an acquired response should not be treated as a universally valid rule. Its domain of validity depends on the conditions under which it was formed and on the new information available.

3.4. Operant Conditioning and the Law of Effect

Operant conditioning, developed notably by B. F. Skinner from Edward Thorndike’s earlier work, concerns the relationships between an organism’s actions and their consequences.

Unlike classical conditioning, in which a stimulus acquires predictive value, operant conditioning examines how consequences modify the likelihood that an action will be repeated.

Thorndike’s law of effect states that responses followed by satisfying consequences tend, under certain conditions, to be strengthened, whereas those leading to unfavourable consequences may become less frequent.

Positive reinforcement consists of presenting a consequence that increases the likelihood of a behaviour. Negative reinforcement consists of removing or avoiding an aversive stimulus, which also increases the likelihood of the behaviour concerned.

Punishment, by contrast, refers to a consequence intended to reduce the likelihood of a response. It must not be confused with negative reinforcement.

This terminological distinction is essential. The positive or negative character of reinforcement does not refer to a moral judgement, but to whether an event is added to or removed from the situation under investigation.

Reinforcement must also be defined by its observable effect. A consequence assumed to be pleasant is not necessarily a reinforcer if it does not actually increase the likelihood of the behaviour.

3.5. Reinforcement Schedules

Skinner’s research showed that the frequency and regularity of reinforcement have a strong influence on acquired behaviours.

Continuous reinforcement occurs whenever the expected response is produced. Intermittent schedules reinforce only some responses.

Some schedules are based on the number of responses emitted; others depend on the passage of time. They may be fixed or variable.

These differences affect the frequency of behaviours, their temporal distribution and their resistance to extinction.

A behaviour reinforced intermittently may persist for a long time when the expected consequence becomes rare. This phenomenon helps explain certain persistent habits, although it cannot by itself account for complex human behaviour.

In artificial intelligence, reinforcement-learning agents use reward signals to evaluate their actions. Rewards may be immediate or delayed, frequent or rare, deterministic or uncertain.

The choice of reward function is decisive. A system optimises the criteria it is actually given, which may not perfectly correspond to the human objective it is intended to serve.

This divergence is particularly important when a reward is used as a substitute for a more complex goal. An agent may exploit a loophole in the problem definition and obtain a high reward without exhibiting the behaviour that was genuinely intended.

3.6. Escape, Avoidance and Fear Learning

Escape and avoidance behaviours are important examples of instrumental conditioning.

In an escape situation, an action terminates an aversive stimulus that is already present. In an avoidance situation, an action prevents or delays an unpleasant event announced by a signal.

These behaviours can become highly stable when the consequences of inaction are strongly negative.

Avoidance presents a particular difficulty: once the action has been acquired, the absence of the feared event may prevent the subject from discovering that the danger has disappeared. The behaviour then continues to be maintained by a reduction in uncertainty or aversive anticipation, even when objective circumstances have changed.

This dynamic has been studied in several models of anxiety and fear-related behaviour. The actual mechanisms of these disorders are nevertheless multifactorial and cannot be reduced to elementary conditioning.

The question is of broader relevance to autonomous systems. An agent that systematically favours an avoidance strategy may reduce certain risks in the short term while losing the opportunity to explore its environment and discover that its estimates were incorrect.

Designing an adaptive system therefore requires a balance between exploration, exploitation of acquired knowledge and risk management.

3.7. The Limits of Conditioning as a General Theory of Behaviour

Conditioning has made a profound contribution to the development of experimental psychology. It has enabled researchers to investigate relationships between events, actions and consequences under controlled conditions.

It is not, however, an exhaustive explanation of human activity.

Complex behaviour also involves representations, goals, explicit knowledge, planning abilities, social norms and reasoning processes.

Observational learning, language comprehension, imitation, problem-solving and concept formation cannot be entirely reduced to the simple repetition of a reinforced response.

The importance of conditioning therefore depends on the phenomenon being studied. It provides a powerful framework for certain forms of learning, but must be combined with other approaches when explaining higher cognitive abilities.

The same caution applies to artificial intelligence. Reinforcement learning is a major method, but it represents neither the entirety of machine learning nor a complete theory of intelligence.

4. Behaviour: From Objective Observation to Scientific Explanation

4.1. Establishing Behaviour as an Object of Scientific Study

Behaviour refers to the observable manifestations of an organism’s activity within its environment.

Scientific psychology gradually developed this concept to distinguish observable events from the interpretations that may be proposed to explain them.

Observing an individual extend an arm, grasp an object and move it is a description of a sequence of actions. Claiming that the person acts out of curiosity, necessity or intention introduces an interpretation that requires additional evidence.

This distinction does not mean that intentions, emotions or representations do not exist. It simply requires direct observations to be distinguished from explanatory hypotheses.

Behavioural measurement therefore relies on defined criteria: frequency, duration, magnitude, latency, accuracy or the probability that a response will occur in a given situation.

These indicators make it possible to compare performance, test hypotheses and determine whether an intervention actually changes the behaviours being observed.

4.2. Watson and Behaviourism

In 1913, John B. Watson published a seminal text of behaviourism in which he advocated a psychology based on the objective observation of relationships between stimuli and responses.

This approach sought to make scientific findings less dependent on introspection and subjective descriptions that are difficult to verify.

Behaviourism helped develop rigorous experimental methods and establish the role of learning in the formation of many behaviours.

Its strictest versions nevertheless encountered limitations when attempting to explain mental representations, language, complex memory and planning.

Cognitive science developed partly in response to these difficulties. It reintroduced the study of internal processes by treating them as objects of modelling and experimentation rather than as merely introspective data.

It is therefore important to distinguish the objective observation of behaviour, which remains essential, from behaviourism, which represents a particular family of theoretical positions.

4.3. Behavioural Units and Parameters

Scientific analysis requires a precise definition of what will be observed.

A behavioural unit may correspond to an elementary movement, a sequence of actions or an objectively identifiable outcome.

In animals, it may refer to exploring an area, searching for food or building a nest. In human beings, it may correspond to a verbal response, a decision or the execution of a task.

The parameters selected depend on the purpose of the study. Frequency makes it possible to count occurrences. Duration measures how long an activity continues. Latency refers to the interval between a reference event and the beginning of a response. Magnitude describes the intensity of a manifestation when it can be meaningfully quantified.

Analysis must also take context into account. An apparently identical behaviour may serve different functions under different circumstances, while several distinct motor sequences may produce an equivalent outcome.

This difficulty is particularly apparent in complex environments. The same gesture may correspond to different intentions, and an action that is appropriate in one context may become dangerous in another.

4.4. Behaviour, Cognition and the Environment

Observable behaviour provides an empirical starting point, but explaining it often requires consideration of the interactions between an organism and its environment.

A course of action may depend on learning history, physiological state, expectations, risk perception and the goals being pursued.

Contemporary sciences therefore investigate behaviour at several levels: biological, cognitive, individual, social and cultural.

Ethology examines animal behaviour in natural and experimental settings. Cognitive psychology studies the processes associated with perception, memory, reasoning and decision-making. Social psychology analyses the influence of groups and interactions. Sociology and anthropology situate behaviour within historical and cultural structures.

These levels are not mutually exclusive. A satisfactory explanation may require simultaneous consideration of individual mechanisms, material constraints and collective norms.

Artificial intelligence encounters a comparable difficulty when a system must operate in the real world. An action cannot be evaluated solely by its physical form: its relevance depends on the state of the environment, the task’s objectives and the consequences it produces.

4.5. Cultural Behaviour and Social Learning

Human behaviour does not develop exclusively through individual experience. It is also shaped by language, institutions, norms, technologies and collective practices.

Anthropology and sociology have shown that ways of acting, eating habits, forms of education and social representations vary across historical and cultural contexts.

Culturalist approaches have helped emphasise the role of social transmission, although some have been criticised for reducing society to a set of behaviours and values without sufficiently analysing power relations, inequalities and material conditions.

Cultural behaviour must be studied in relation to social organisation, individual trajectories and historical transformations.

Social learning is also highly relevant to robotics and artificial intelligence. An agent can acquire information by observing another agent’s actions, exploiting demonstrations or receiving instructions.

Imitation learning can, for example, train a robot to reproduce observed trajectories or sequences of actions. However, reproducing a demonstration does not guarantee that the system understands the action’s objectives or can adapt it to a different situation.

The challenge is to transform observation into a sufficiently robust ability to act beyond the exact conditions of the demonstration.

5. Memory Mechanisms in Artificial Intelligence

5.1. Computer Memory Is Not Human Memory

The term memory has several meanings in computing.

Random-access memory (RAM) temporarily holds the data needed to execute programs. Persistent storage retains information after a process has stopped. Databases organise information into structures designed to facilitate consultation, modification and retrieval.

In artificial intelligence systems, memory may also refer to parameters learned during training, information present in the context of a query, documents retrieved from an external database or an agent’s interaction history.

These forms of computer memory do not serve the same purposes.

The parameters of a neural network encode regularities learned during training. They do not necessarily constitute an explicit archive of the examples used. A document-retrieval system, by contrast, can retain identifiable texts and search for them in response to a query.

A language model’s context allows it to process a set of information during an interaction. Depending on the system’s architecture, that information may not be retained after the session ends.

It is therefore necessary to distinguish physical storage, learned representations, temporary context and an agent’s persistent memory.

5.2. Neural Network Parameters as Distributed Memory

Artificial neural networks learn from data by adjusting numerical parameters.

During training, an algorithm modifies these parameters to improve results according to an objective function. Relevant regularities can then be represented in a distributed manner through the relationships among the network’s different units.

This organisation makes it possible to generate responses without systematically consulting every training example.

It nevertheless has limitations. Learned information may be difficult to isolate, update or remove without affecting other capabilities. Parameters do not always provide an explicit representation of the origin of a piece of knowledge, and a model may produce incorrect results despite extensive training.

Distributed memory is therefore a functional property of artificial architectures, not evidence of equivalence with biological mechanisms.

5.3. The Context Window of Language Models

Large language models process text sequences through a context window whose size depends on the architecture and configuration of the system.

This window allows the model to take into account information contained in a conversation, document or instruction.

It should not be confused with lasting personal memory. Information included in the context can be used during processing without automatically being retained for a subsequent interaction.

Nor does a larger context guarantee that every detail will be used correctly. The position of information, redundancy, ambiguity and task complexity all influence performance.

Architectures capable of processing long sequences seek to improve the use of information located far apart within a document, but simply increasing the context window does not solve every retrieval and reasoning difficulty.

5.4. Retrieval-Augmented Generation

Retrieval-augmented generation, generally referred to by the acronym RAG, combines a generative model with a document-retrieval mechanism.

The principle is to search an external database for relevant information and then provide it to the model when generating a response.

This architecture can improve access to specialised documents, up-to-date information or knowledge specific to an organisation.

It also makes it possible to distinguish more clearly between document storage and language processing. Information can be updated in the database without necessarily requiring the model to undergo a complete retraining process.

However, the quality of a RAG system depends on several factors: document quality, text segmentation, indexing methods, the relevance of retrieved results and the model’s ability to use the supplied material correctly.

An erroneous retrieval can produce a misleading response. A correct source may also be misinterpreted or insufficiently taken into account.

RAG therefore does not eliminate the risk of error; it provides an architecture that can link certain responses more closely to retrievable information.

5.5. Memory in Autonomous Agents

Artificial intelligence agents can combine several forms of memory to maintain operational continuity across different stages of a task.

Working memory retains the elements needed for immediate action. Artificial episodic memory may record previous interactions or events. Semantic memory may organise general knowledge or facts extracted from documents. Procedural memory may store reusable instructions, strategies or procedures.

These labels describe functional roles. They do not demonstrate that an agent possesses the same cognitive systems as a human being.

The challenge is nevertheless significant: a system carrying out a multi-stage task must retain relevant decisions, avoid unnecessarily repeating operations and revise its plan when new information emerges.

Effective memory requires a selection policy. Retaining everything can increase storage costs, complicate retrieval and encourage the use of outdated information. Conversely, an overly restrictive memory may eliminate elements essential to coherent action.

The design of autonomous systems must therefore balance retention, compression, retrieval, verification and updating.

6. Machine Learning: Fundamental Principles

6.1. Supervised Learning

Supervised learning involves training a model on examples associated with expected outcomes.

A classification system may receive labelled images to learn to distinguish several categories. A forecasting model may use historical data to estimate a future value. A language system may be trained on examples of texts and responses corresponding to a particular task.

The model adjusts its parameters to reduce a loss function that quantifies the discrepancy between predictions and target values.

Learning quality depends notably on the representativeness of the data, the definition of the task, the model architecture and the evaluation method.

Biased or insufficient data may produce a model that performs well under certain conditions but proves unreliable under others.

Supervised learning does not directly correspond to classical or operant conditioning. It is a computational method whose training signal has a specific structure.

6.2. Unsupervised and Self-Supervised Learning

Unsupervised learning seeks to extract structures from data without consistently relying on explicit labels supplied by a human annotator.

It can be used, in particular, to identify clusters, discover compact representations or analyse regularities within a dataset.

Self-supervised learning, meanwhile, constructs training objectives from the data themselves. Language models use, among other approaches, token-prediction tasks or the reconstruction of masked information, depending on the architecture and method employed.

These approaches make it possible to exploit large volumes of data, but they do not guarantee that the learned structures correspond directly to the categories or explanations sought by users.

A model may identify useful regularities without possessing an explicitly verified representation of the causal relationships that produce them.

The distinction between prediction, understanding and knowledge therefore remains essential when interpreting the capabilities of modern systems.

6.3. Reinforcement Learning

Reinforcement learning examines how an agent selects actions to maximise cumulative reward within an environment.

An agent observes a state, selects an action, receives a new observation and, potentially, a reward. It must learn an action policy that takes both immediate and future consequences into account.

Unlike conventional supervised learning, the correct action is not necessarily indicated at every stage. The agent often has to explore different possibilities to identify those that produce the most favourable outcomes.

This method is used in fields such as control, certain games, strategy optimisation and robotics.

It nevertheless presents several challenges: costly exploration, sparse rewards, training instability, dependence on simulation and difficulty transferring a policy to a real-world environment.

A poorly defined reward can lead to undesirable behaviour. The agent’s quality therefore depends as much on how the problem is formulated as on the algorithm being used.

6.4. The Problem of Catastrophic Forgetting

Continual learning aims to enable a model to incorporate new data or tasks without excessively losing previously acquired capabilities.

Neural networks may encounter a problem known as catastrophic forgetting: adjusting parameters to a new task can substantially degrade performance on earlier tasks.

At a functional level, this phenomenon recalls certain difficulties biological systems encounter when learning competing information. It does not mean that the underlying mechanisms are identical.

Strategies for limiting forgetting include retaining old examples, protecting selected parameters, applying regularisation, separating modules and using external memory systems.

The choice of method depends on data availability, computational constraints, privacy requirements and the nature of the tasks.

This issue is particularly important for agents that must evolve in changing environments. A system that remains useful over time must be able to incorporate new information while retaining the skills it still needs.

6.5. Transfer Learning

Transfer learning involves reusing knowledge or representations acquired on one task to improve performance on another.

A model pretrained on a large dataset can be adapted to a specialised task through additional training or a limited modification of its parameters.

Compared with training an entirely new model, this strategy often reduces data and computational requirements.

However, it may fail when the source and target tasks differ too greatly or when features that were useful in the original domain become misleading in the new context.

Transfer learning therefore connects with a central question in cognitive science: under what conditions does a previous acquisition facilitate a new skill?

The answer depends on the structure of the tasks, the representations available and the way in which knowledge is mobilised.

7. From Memory to Action: Integrated Artificial Intelligence Models

7.1. Perceiving, Learning, Remembering and Acting

An autonomous system generally has to perform several complementary operations.

It must perceive its environment, construct a usable representation of the situation, retain relevant information, select an action and evaluate its consequences.

These operations do not necessarily form a strictly linear sequence. New observations may alter the interpretation of previous data, and actions may be used to obtain additional information.

In a mobile robot, visual perception may help locate obstacles. Memory retains a representation of the environment. Planning determines a route. The control system executes movements and adjusts the trajectory when conditions change.

Learning can improve each of these stages, but not all of them require the same mechanism. One part of the system may rely on explicit rules, another on statistical models and a third on control methods.

Operational artificial intelligence therefore often depends on integrating several methods rather than using a single algorithm.

7.2. Feedback Loops

Feedback refers to a mechanism through which the results of an action are reused to modify a system’s operation.

In a control system, a measured quantity is compared with a reference value. The observed discrepancy is used to adjust the action in order to reduce the error.

Feedback loops are fundamental to automatic control, robotics and many industrial systems.

They can be combined with learned models. A controller may, for example, use an estimate of the system’s state to adjust its control signal, while a learning algorithm gradually improves that estimate.

However, feedback must be distinguished from learning. A system may continually correct its actions without permanently modifying its parameters. Conversely, a model may learn from a dataset without directly interacting with the system that produced it.

This distinction links the behavioural sciences with engineering: the consequences of an action can be used both to regulate a system immediately and to improve its future capabilities.

7.3. Learning in Robotics

Robotics is a particularly useful field for studying the integration of perception, memory, learning and behaviour.

A robot generally has to account for physical constraints: inertia, friction, sensor uncertainty, environmental variations and actuator limitations.

Learned models can contribute to object recognition, motion estimation, manipulation and navigation.

Imitation learning makes it possible to reuse demonstrations. Reinforcement learning can optimise certain action strategies. Planning and control methods translate objectives into executable commands.

The most robust systems often combine these approaches with safety mechanisms, operational limits and verification procedures.

The ability to perform an action effectively in simulation does not guarantee that it will be effective in the real world. Differences between simulation and reality, hardware failures and unforeseen events may compromise an autonomous system’s behaviour.

Learning must therefore be evaluated in terms of stability, safety and adaptability, not merely average performance on a test set.

8. The Limits of Analogies Between the Brain and Artificial Intelligence

8.1. Different Architectures

The human brain and artificial neural networks share certain abstract ideas, notably the use of numerous interconnected units and the modification of parameters through experience.

This resemblance is not sufficient to establish biological equivalence.

The brain operates through neurons, synapses, neurotransmitters, electrochemical processes, glial cells and networks subject to complex physiological constraints.

Artificial networks use mathematical operations executed on computer hardware. Even when inspired by neurobiological principles, their learning mechanisms, temporal dynamics and constraints differ considerably from those of living organisms.

The comparison is therefore meaningful when it concerns a clearly defined abstract function, such as generalisation or learning from errors. It becomes misleading when it assumes that two systems producing similar results must possess the same internal mechanism.

8.2. Memory Does Not Guarantee Understanding

A system may retain or reproduce information without being able to assess its significance correctly.

Language models can produce convincing answers that nevertheless contain factual errors. They may exploit statistical regularities without independently verifying every claim they generate.

Adding external memory can improve access to documents, but it does not guarantee their correct interpretation.

Reliability therefore requires appropriate validation mechanisms: consulting sources, running tests, checking calculations, assessing consistency and, when the consequences are significant, obtaining human oversight.

It is also necessary to distinguish the availability of information, its relevance to the task and the justification for using it.

8.3. Reward Is Not Equivalent to Emotion

A reinforcement-learning agent can optimise a numerical reward. This operation does not demonstrate that it experiences pleasure, pain or subjective motivation.

In living organisms, affective states are embedded in biological, bodily, cognitive and social mechanisms. Computational rewards are variables defined within an optimisation procedure.

A functional analogy may be useful for designing a system that favours certain actions, but it does not justify concluding that subjective experience is present.

This distinction is particularly important when a system uses emotional language or adopts socially familiar behaviours. The form of its response is not sufficient to establish the nature of its internal states.

8.4. Biases, Errors and Robustness

Learning depends on data, objectives and evaluation conditions.

A model trained on historical data may reproduce inequalities present in those data. A recognition system may function correctly in one environment and deteriorate in another. An autonomous agent may favour a strategy that maximises its performance indicator while producing undesirable consequences.

These difficulties remind us that adaptation does not automatically mean progress.

A system can efficiently learn an unsuitable rule, reinforce a dangerous strategy or incorrectly generalise a statistical relationship.

Artificial intelligence design must therefore include bias assessment, uncertainty measurement, the detection of unusual situations and the ability to interrupt or correct an action.

Learning does not replace responsible system design. It is one component of it.

9. Contemporary Applications: Education, Healthcare, Industry and Human Interaction

9.1. Education and Personalised Learning

Artificial intelligence systems can help tailor exercises to a learner’s observed level, detect certain recurring errors and suggest revision activities.

Spaced repetition and active recall methods can be integrated into educational tools that adjust the timing of exercises according to results.

These systems do not replace an understanding of educational objectives, human support or the comprehensive assessment of a student’s development.

Performance measured by an application does not necessarily reflect every dimension of learning. Motivation, prior knowledge, social context and the ability to transfer acquired skills to new situations must also be taken into account.

A reliable educational AI system must therefore provide useful indications without confusing observed scores with a comprehensive measure of a learner’s abilities.

9.2. Healthcare and Biomedical Research

Artificial intelligence can help analyse biomedical data, identify regularities and support certain research tasks or clinical decisions.

Applications may include image analysis, medical-record processing, risk modelling and the investigation of biological mechanisms.

These uses require validation appropriate to the population, clinical context and intended purpose of the tool.

A model that performs well on retrospective data may become less reliable when deployed in another institution or used with a different population.

Biological memory, learning and behaviour also remain fundamental research topics for understanding cognitive disorders, neurodegenerative diseases and the mechanisms of rehabilitation.

AI can contribute to this research, but findings must be interpreted in the light of experimental data and available biological knowledge, without turning a calculated correlation into an unproven causal explanation.

9.3. Industry and Predictive Maintenance

In industry, machine-learning systems can analyse sensor measurements to detect anomalies or estimate the risk of equipment failure.

These systems learn regularities from historical data and can improve the monitoring of industrial installations.

However, they must be adapted to changes in machinery, operating conditions and faults that have never been observed during training.

Predictive maintenance therefore relies on a combination of data, physical knowledge, safety rules and intervention procedures.

The ability to recognise an anomaly does not guarantee that its cause has been identified. A reliable system must distinguish between detecting an unusual signal and explaining the mechanism that produced it.

9.4. Human–Machine Interaction

Conversational assistants and software agents use language-processing mechanisms to interpret requests, exploit context and generate responses.

Conversational memory can improve continuity when it retains relevant information and keeps it correctly updated.

Nevertheless, it raises questions concerning privacy, control and reliability. Retained information may become outdated, be interpreted incorrectly or be used in an inappropriate context.

Systems must therefore distinguish temporary information, enduring preferences and data that require verification.

The quality of an interaction depends not only on a system’s ability to respond, but also on its capacity to recognise the limits of its knowledge, request clarification and correct its mistakes.

10. Towards an Integrated Understanding of Memory, Learning and Behaviour

10.1. A Functional Chain Rather Than a Sequence of Isolated Faculties

Memory, learning, conditioning and behaviour can be understood as interdependent components of an adaptive system.

Experience provides information. Certain events modify the available representations. Learning transforms abilities or dispositions to act. Memory makes it possible to retain and retrieve some of these acquisitions. Behaviour produces actions whose consequences influence subsequent experiences.

This organisation forms a loop: actions modify the environment, new observations provide information about their consequences, and acquired knowledge influences future decisions.

In living organisms, this dynamic involves multiple biological and cognitive processes. In artificial systems, it can be implemented through perception, memory, planning, control and learning modules.

The value of this integrated perspective lies in moving beyond explanations that artificially isolate a single mechanism. Observed behaviour may depend on an old habit, recently acquired information, an environmental constraint and an immediate objective.

Scientific analysis must therefore determine how these factors interact rather than seek a single cause for every response.

10.2. Learning Without Forgetting: A Shared Problem, Distinct Solutions

Living organisms and artificial systems must deal with new information, limited resources and potentially competing acquisitions.

In the life sciences, consolidation, memory reactivation and plasticity mechanisms contribute to the stability and transformation of acquired knowledge.

In artificial intelligence, external memory, retraining strategies, regularisation and continual learning seek to maintain a balance between acquiring and retaining information.

These approaches address a common functional difficulty, but their mechanisms must not be confused.

Forgetting can be a drawback when essential information becomes inaccessible. It may also help limit the influence of irrelevant details or adapt responses to a changing environment.

Designing an effective memory system therefore does not necessarily mean retaining everything. It requires determining which information must remain accessible, for how long and under what conditions it should be revised.

10.3. The Role of the Environment in Intelligence

Intelligence does not manifest itself independently of the conditions in which a system operates.

Living organisms develop their abilities within sensory, social and material environments that influence perception, learning and the formation of habits.

Artificial systems likewise depend on their training data, available interfaces, computational constraints and deployment conditions.

A capability demonstrated in a controlled environment does not automatically transfer to every situation. A change in context may alter the available information, the relevance of strategies and the nature of possible errors.

The evaluation of artificial intelligence must therefore go beyond performance on isolated tests. It must examine robustness, generalisation, the ability to manage uncertainty and adaptation to change.

This requirement brings the sciences of learning closer to engineering: the value of a model also depends on its ability to operate reliably under the conditions for which it was designed.

10.4. Behaviour as the Observable Outcome of Adaptation

Behaviour represents the level at which the consequences of learning become directly observable.

It makes it possible to determine whether an acquisition actually modifies actions, whether a strategy withstands changes in context and whether a system can correct its errors.

However, behaviour alone does not always reveal the mechanism that produced it.

Two systems may provide the same response by relying on different processes. Conversely, the same mechanism may produce different behaviours depending on environmental conditions.

This distinction requires an empirical approach based on controlled experiments, comparisons and appropriate measurements.

In artificial intelligence, behavioural tests can assess performance, but, whenever possible, they should be supplemented by analysis of the data, parameters, decision mechanisms and conditions of use.

The objective is not merely to establish that a system works, but to determine under what circumstances, for what reasons and within what limits it works.

Conclusion: From Biological Memory to Adaptive Artificial Intelligence

Memory, learning, conditioning and behaviour are four essential dimensions in the study of systems capable of modifying their responses in accordance with experience.

Their scientific history reveals a gradual transition from predominantly philosophical conceptions to experimental approaches based on observation, measurement and modelling. Experimental psychology made it possible to characterise the effects of repetition, time, reinforcement and interference. Neuroscience highlighted the role of synaptic plasticity, consolidation and the organisation of brain networks. Cognitive science enriched these models by investigating representations, working memory, reasoning and generalisation.

Artificial intelligence revisits some of these questions in computational form. Machine learning makes it possible to extract regularities from data. Reinforcement learning investigates the optimisation of actions according to their consequences. Artificial memory systems make it possible to retain information, retrieve documents and maintain operational continuity. Robotics brings these capabilities together in devices confronted with the constraints of the physical world.

This convergence does not mean that machines reproduce the biological mechanisms of memory or necessarily possess the mental states associated with human behaviour. Rather, it indicates that several disciplines address functionally comparable problems: retaining useful information, learning from experience, selecting an action and adapting to a changing environment.

The distinction between functional similarity and identity of mechanisms is essential to a rigorous understanding of AI. It makes it possible to draw on the lessons of biology and psychology without attributing properties to artificial systems that have not been demonstrated.

The contemporary challenge is now to design systems capable of learning without unnecessarily losing previous acquisitions, remembering without accumulating obsolete information, generalising without blindly reproducing biases and acting without exceeding the limits of their reliability.

Memory provides the resources of the past, learning transforms capabilities, conditioning sheds light on certain relationships between events and consequences, while behaviour makes the effects of these processes observable. Their interaction constitutes an essential framework for understanding biological intelligence, designing adaptive artificial systems and accurately assessing both their possibilities and their limitations.

FAQ: Memory, Learning, Conditioning and Artificial Intelligence

What is the difference between memory and learning?

Memory refers to the processes that enable information to be encoded, retained and retrieved. Learning involves relatively lasting changes in knowledge, skills or dispositions to act as a result of experience. The two processes are closely related, but changes in ability do not necessarily require conscious recollection, and memorised information is not always immediately accessible.

What is the difference between classical and operant conditioning?

Classical conditioning is based on learning predictive relationships between events. A stimulus may acquire the ability to elicit a response because it regularly signals another event. Operant conditioning concerns the effects of consequences on the likelihood that an action will be repeated. These two forms of learning can interact within the same situation.

How does artificial intelligence memory work?

AI memory may refer to learned parameters, information contained in a context window, an external document database or an interaction history retained by an agent. These mechanisms serve different functions. Parameters encode learned regularities, whereas an external database can retain searchable documents. None of these mechanisms should automatically be equated with human autobiographical memory.

Can artificial intelligence learn like a human being?

AI can acquire capabilities from data, correct certain errors, generalise regularities and adapt its actions. However, its learning mechanisms differ from those of biological organisms. The similarities mainly concern certain abstract functions, not a demonstrated identity between cognitive or neural processes.

What is reinforcement learning?

Reinforcement learning is a machine-learning method in which an agent adapts its actions according to the rewards it receives within an environment. It seeks to improve an action policy by taking immediate and future consequences into account. The quality of the result depends notably on how the reward is defined, the opportunities for exploration and the evaluation conditions.

Why can an AI forget what it has learned?

When a neural network is trained on new data, modifying its parameters may degrade its performance on earlier tasks. This phenomenon, known as catastrophic forgetting, is a major challenge in continual learning. Regularisation, example retention and external-memory methods can limit certain losses, but do not systematically eliminate them.

Why is repetition not always enough for learning?

Repetition can strengthen an acquisition, but its effectiveness depends on attention, active recall, the spacing of practice sessions, feedback and prior knowledge. Overly uniform practice may create an impression of mastery without guaranteeing durable retrieval. Learning therefore requires an approach suited to its objectives and to the conditions under which knowledge will be used.

What role does behaviour play in evaluating AI?

Observable behaviour makes it possible to measure a system’s responses under defined conditions. It helps assess accuracy, robustness, adaptability and the effects of changes to training. However, a correct response does not necessarily reveal the mechanism that produced it. Evaluation should therefore combine behavioural tests, technical analysis and checks appropriate to the context.

What is the connection between memory, learning and robotics?

A robot must retain relevant information, use its observations, learn from certain experiences and adjust its actions according to its environment. Memory contributes to continuity of action, learning improves certain capabilities, and control mechanisms regulate execution. Integrating these elements makes it possible to build more adaptive systems, provided that physical constraints and safety requirements are taken into account.

Why is it important to distinguish artificial intelligence from biological intelligence?

Both fields investigate certain common questions, including learning, memory, prediction and adaptation. However, biological systems rely on cellular, physiological and cognitive mechanisms that artificial architectures do not reproduce identically. This distinction prevents observable performance from being confused with evidence of subjective experience or an identical mental mechanism.