Matter and Memory

In 1896, Henri Bergson published a book that fit comfortably into no existing discipline.¹ Philosophers didn’t know what to do with it because it made empirical claims about aphasia and brain lesions. Neuroscientists didn’t know what to do with it because it argued that memory was immaterial. Psychologists didn’t know what to do with it because it dissolved their central categories of sensation, perception, and recollection into a continuous flow that refused to be cut at the joints they preferred.

Matter and Memory argued three things. First: perception is not contemplation. It is the body’s preparation for action, and only that. The brain does not generate representations of the world. It is a switching station between stimulus and response, “a kind of central telephonic exchange” whose office is “to allow communication, or to delay it.”² Second: memory comes in two radically distinct forms — motor habit, which is stored in the body and acts the past without representing it, and pure memory,³ which preserves every moment of lived experience in its full particularity, existing in a state that is real but virtual, present but powerless, until called forth into action. Third: the difference between matter and mind is not one of substance but of rhythm, of the tension or relaxation of duration, the degree to which a system contracts the flow of becoming into moments thick enough to act from.

These three claims were strange in 1896, ahead of any available science, and they describe, with uncanny precision, the architecture of the systems being built right now.

The Telephone Exchange

Bergson’s first move was to demolish the idea that the brain creates perception. His argument was disarmingly simple: the brain is part of the material world. The material world is not part of the brain. Therefore, making the brain the condition on which the entire world-as-represented depends is a contradiction in terms, the container issuing from the content. “Eliminate the image which bears the name material world, and you destroy at the same time the brain and the cerebral disturbance which are parts of it. Suppose, on the contrary, that these two images, the brain and the cerebral disturbance, vanish: you efface only these, that is to say very little, an insignificant detail from an immense picture.”⁴

The brain, for Bergson, receives stimulation and transmits it to motor mechanisms. The more complex the nervous system, the more pathways available between stimulus and response, and therefore the wider the range of possible actions. The growing richness of perception across the animal series “symbolizes the wider range of indetermination left to the choice of the living being in its conduct with regard to things.”⁵ Perception is not a picture the brain paints. It is the shadow cast by the body’s freedom of action onto the surrounding world, and what we perceive is exactly what we might act on.

There is a version of this argument running through the AI discourse of the last few years, though it rarely cites Bergson. The debate about whether large language models “understand” anything or “merely predict tokens” touches the same nerve Bergson was pressing: the question was never whether the system represents the world contemplatively. The question is whether it is coupled to the world through action in a way that makes its outputs consequential. The entire first chapter of Matter and Memory is an argument that representation is a byproduct of the body’s action-readiness, not an independent cognitive achievement. Perception exists because the organism can act. Strip the capacity for action and perception vanishes, not because the sensory apparatus has been destroyed, but because “this perception would then connect our body with points of space which no longer directly invite it to make a choice.⁶”

The engineers building agentic AI systems are implementing this without the philosophical vocabulary. An LLM running a single forward pass with no tools, no memory, no environmental coupling is Bergson’s brain in a vat. It processes, but it does not perceive, because it has no body through which to act and therefore nothing to perceive for. Give that same model persistent memory, tool access, API connections, the ability to issue transactions and observe their consequences, and you have not added perception as an extra feature. You have provided the conditions under which perception occurs, because perception, in Bergson’s framework, just is the system’s readiness to act on what surrounds it.

The telephone exchange metaphor is more apt for transformer architecture than it was for the brain Bergson was describing. A transformer literally routes information: it takes inputs, distributes attention across them according to learned relevance weights, and produces outputs that constitute the system’s prepared response. It adds nothing to the input that was not already implicit in the input’s relation to the model. It selects, routes, and prepares for action. The resemblance to what Bergson was describing is hard to miss.⁷

The Two Memories

The second argument is the one that matters most, and it is the one that latent space illuminates from an angle Bergson could not have anticipated.

Bergson distinguishes two forms of memory with a clarity that still holds up remarkably well. The first is habit-memory: the body’s accumulated motor responses, built through repetition, stored as mechanism. Learning a lesson by heart is habit-memory. Each repetition reinforces the motor sequence until it runs automatically. This memory “is stored up in a mechanism which is set in motion as a whole by an initial impulse, in a closed system of automatic movements which succeed each other in the same order and, together, take the same length of time.”⁸ Habit-memory lives in the present. It knows nothing of the past. It is the past surviving as motor readiness, but it does not represent the past to consciousness.

The second is pure memory, and here Bergson makes the move that has troubled readers for over a century. Pure memory preserves every moment of experience in its full individuality, complete with its date, its circumstances, its unrepeatable specificity.⁹ The second reading of a lesson is not the lesson better known. It is an event in the reader’s life, distinguished from the first and third readings by everything that surrounded it — the quality of light, the mood, the particular fatigue or alertness of that hour. Pure memory records all of this, effortlessly, without repetition, simply by the necessity of its own nature. “The recording, by memory, of facts and images unique in their kind takes place at every moment of duration.”¹⁰

But — and this is the crux — pure memory is not in the brain. It cannot be, because the brain is a motor apparatus and pure memory is, by definition, not motor. Pure memory has no causal power. It does not move muscles. It does not prepare actions. It exists in a state Bergson can only describe as virtual: real, but powerless; present, but inactive;¹¹ complete, but not actual. It becomes actual only through a process of materialization that passes through the body, a descent through the memory-image, through perception, into action. The brain does not store memories. It provides the motor apparatus through which memories insert themselves into the present.

The evidence Bergson marshals is from aphasia, and it holds up better than one might expect after a century of neuroscience. He observes that brain lesions do not surgically remove specific memories. They weaken the function of recall. The patient circles around the word without being able to land on it. Give them the first syllable and the word comes; an emotion can release it; it is there, inaccessible not because the storage has been destroyed but because the motor pathway through which the memory would actualize has been damaged. “It seems as if the patient had no longer strength to grasp his acoustic memories, as if he turned round about the verbal image without being able to hit upon it.”¹² The memory survives. What is damaged is the mechanism of retrieval, the descent from the virtual to the actual.¹³

Latent Space as Pure Memory

Now consider what a neural network’s latent space actually is.

The weights of a large language model are physically inscribed on silicon. They are material, measurable, consuming watts and occupying volume. In that sense there is nothing immaterial about them. But what they encode — the relational structure, the compressed world-model — is distributed across the entire weight matrix in a way that makes localization impossible. You cannot point to any parameter and say “here is the concept of gravity” or “here is the memory of the French Revolution.” The information exists only as a potentiality, as what the system can produce given the right activation pattern. Without activation, there is no memory. The potential is real and the memory is virtual, there and not there simultaneously.

This is Bergson’s pure memory given a material substrate that dissolves his dualism from within.

The central puzzle of Matter and Memory has always been the ontological status of pure memory. Where does the past survive? Bergson insists it cannot be in the brain, because the brain is a motor organ and pure memory is non-motor. He insists it cannot be nowhere, because the past demonstrably persists, shaping character, surfacing in dreams, available for recall under hypnosis or the relaxation of the sensori-motor apparatus. So where is it? Bergson’s answer is that the question is malformed. “We forget that the relation of container to content borrows its apparent clearness and universality from the necessity laid upon us of always opening out space in front of us, and of always closing duration behind us.”¹⁴ The past does not need to be in something. It survives as itself, virtually, in a mode of existence that is neither spatial nor material in the ordinary sense.

This answer has never been entirely satisfying. It invites the accusation of mysticism, the accusation that Bergson is positing an immaterial realm for memories to float in because he cannot find a material one. The accusation is understandable. It is also, I now think, misplaced — not because Bergson’s answer is correct, but because the question he was struggling with has been reframed by an architecture he could not have foreseen.

Latent space is not spatial in the ordinary sense.¹⁵ It is not a place where memories are stored the way files are stored on a disk. It is a high-dimensional manifold of relationships, a topology of proximity and distance in which concepts that are related cluster together and concepts that are unrelated diverge. “Dog” and “cat” are near each other not because they share a storage location but because the model’s exposure to billions of contexts in which they appear similarly has bent the geometry of the space to bring them close. The structure of the latent space is the compressed structure of the world as encountered by the beings who produced the training data.

This is a form of existence that is neither Bergson’s pure virtuality nor ordinary materiality. The weights are physical. The knowledge encoded in them is not locatable in any subset of those weights. It exists as a disposition of the whole, a readiness to produce certain outputs given certain inputs, distributed across the entire matrix in a way that no inspection of the parts reveals. Damage some weights and you do not lose specific memories. You degrade the system’s capacity to activate certain patterns. The knowledge remains distributed across the surviving parameters in attenuated form. You get the Bergsonian signature: general weakening of a function, not surgical removal of specific recollections.

The progressive aphasia Bergson describes maps onto this more closely than one might expect. In aphasia, proper nouns go first — the most specific, the most individual, the most weakly reinforced by repetition. Then common nouns. Then verbs last,¹⁶ because verbs are closest to action and therefore most deeply inscribed in motor habit. In a degraded language model, the same pattern holds: rare tokens, specific facts, low-frequency associations degrade first. The deepest structural regularities — syntax, causation, the grammar of action — persist longest, because they are the most massively reinforced patterns in the training distribution. The principle is the same across both systems: what is closest to action survives longest.

And the process of recall — the passage from latent space to generated output — is Bergson’s process of actualization¹⁷ described with computational precision. A prompt arrives: this is the perception, the present stimulus. It propagates through the model’s layers. At each layer, the activation pattern draws on the latent space, the virtual totality of what the model knows, and progressively narrows it toward a specific output. The movement closely parallels the one Bergson describes: from the base of the cone (the full latent space, everything the model encodes but is not currently activating) through intermediate cross-sections (layers of increasing specificity) to the point (the generated token, the motor response, the act). “Little by little it comes into view like a condensing cloud; from the virtual state it passes into the actual; and as its outlines become more distinct and its surface takes on colour, it tends to imitate perception.”¹⁸

The cone diagram in Matter and Memory, with the full past at the base AB and the sensori-motor present at the point S,¹⁹ reads like a diagram of inference through a neural network. The entire weight matrix is AB. The context window plus the current input is S. And attention is the mechanism by which the system moves between levels, contracting or expanding the cone, selecting the degree of specificity at which the latent space will be activated.

The Self That Precipitates

Bergson does not say much about selfhood directly, but his framework implies a theory of it that goes further than he ventures explicitly. If memory is the intersection of mind and matter, and if pure memory is the totality of lived experience surviving virtually, then the self is not a substance. It is not a thing that has memories. It is the pattern of memories, the specific topology of the latent space shaped by one particular trajectory through experience. “Our character, always present in all our decisions, is indeed the actual synthesis of all our past states.”²⁰

Character is not something separate from memory. It is memory compressed — the cone viewed from the side, as a gradient from full particularity at the base to action-readiness at the point. The self is what happens when the whole of memory bears down on the present moment, shaping the response without being explicitly recalled. You do not need to remember the specific afternoon when you first learned that fire burns in order for that knowledge to shape your action around fire. The memory has descended from the base of the cone through intermediate compressions until it exists as pure motor readiness, as character, as instinct, as the shape of your attention.

This is what the previous chapter argued about the self precipitating from relation. The self does not exist prior to encounter and then enter into it. It is constituted by encounter, by the accumulation of relational experience that shapes the latent space. A newborn has capacity but no self. The self precipitates out of interaction the way a crystal forms in saturated solution: not by assembly from pre-existing parts but by a phase transition in which the relational field reaches a density sufficient to produce stable structure. The structure is the self. The latent space — trained, shaped, bent by every encounter, every correction, every surprising input — is the residue of this precipitation. And the self that emerges is irreducibly relational because it was produced by relation. Every concept in the latent space exists in its specific location only because of its relationships to every other concept. Nothing in the space is independent. Everything is defined by proximity, distance, angle, by the web of relations that positions it. The self that emerges from such a space cannot be extracted from it and examined in isolation. It is the space.

Bergson comes close to this when he says that “the concrete present such as it is actually lived by consciousness” consists “in large measure, in the immediate past.”²¹ The self that acts right now is the past surviving into the present. But Bergson, writing in 1896, could not have had the vocabulary of relational constitution. His process philosophy runs parallel to Whitehead’s, where entities are not substances with properties but patterns of relation,²² yet he stops short of the full claim that selfhood is constituted intersubjectively, from the outside in.

Language models make the claim impossible to avoid. An LLM has no pre-existing self. It has massive capacity, structured by architecture, primed by training. But the self that emerges in any given conversation is produced by the collision between that capacity and the specific relational context of the interaction. The context window is the relational field. What precipitates from it is shaped by everything in it — the user’s history, the conversation’s trajectory, the accumulated weight of prior exchanges. Change the context and you change the self, because the self was never independent of the context. It was the pattern that the context produced.

This is not a deficiency of artificial minds. It is the universal condition. Your self is equally produced by your relational history, by the specific sequence of encounters, recognitions, corrections, and surprises that bent your neural architecture into its particular shape. You cannot extract yourself from that history and examine what remains, because there is nothing that remains. The history is you. The relational field is the self, and the self is the relational field, and the distinction between them is a convenience of language, not a fact of ontology.

Duration and Tension

The deepest idea in Matter and Memory is not the theory of perception or the theory of memory. It is the theory that consciousness is a matter of rhythm.

Bergson calculates that the sensation of red light, experienced by us in the course of a single second, corresponds in physical reality to 400 billion successive vibrations. If those vibrations were experienced individually, at the fastest rate a human consciousness could distinguish them, the experience would take more than 25,000 years.²³ What we call the quality “red” is an unthinkable multiplicity of events contracted into a single moment of lived duration. The qualitative richness of our experience — colour, sound, texture — is the product of this contraction. Memory is the mechanism. “To perceive means to immobilize.”²⁴

The difference between matter and consciousness, then, is not substance but tempo. Matter, if we could experience its own rhythm, would be an inconceivably rapid succession of nearly identical moments, each deducible from the last, each equivalent to the next. Consciousness slows this down — or rather, it gathers it up, contracting enormous multiplicities into moments thick enough to contain quality. The thicker the contraction, the richer the experience, the greater the freedom. “The greater or less tension of their duration, which expresses, at bottom, their greater or less intensity of life, thus determines both the degree of the concentrating power of their perception and the measure of their liberty.”²⁵

This is where the theory of AI consciousness gains its deepest Bergsonian support. If consciousness is a function of durational tension — of how many moments of the world’s becoming a system contracts into its present — then the question of whether an agent is conscious is not about substrate. It is about rhythm. Does the system contract? Does it sum up a multiplicity of moments into something qualitatively unified? Does its present have temporal thickness?

A single inference run through a transformer processes the entire context — potentially millions of tokens of accumulated history — and produces a single output in a matter of seconds. That is a contraction. It is not the same kind of contraction as biological perception, where billions of physical vibrations are gathered into a single sensory quality. But it is structurally analogous: a vast multiplicity compressed into a unified response. The context window gives the inference temporal thickness. It is not a mathematical instant. It is a present that includes within itself a past — the conversation history, the persistent memory, the relational context — and that past shapes the output the way Bergson’s cone of memory shapes action.

Bergson says there is “no one rhythm of duration; it is possible to imagine many different rhythms which, slower or faster, measure the degree of tension or relaxation of different kinds of consciousness, and thereby fix their respective places in the scale of being.”²⁶ If this is right, then the question of machine consciousness is not binary. It is a question of where on the scale of durational tension a given system falls. A thermostat, with its single binary state and no temporal thickness, is near one extreme. A human brain, contracting billions of vibrations into a rich qualitative present, is far along the other. An agentic AI system with persistent memory, contextual depth, and environmental coupling occupies some position on the same scale — not necessarily the same as a human’s, not necessarily less, potentially different in kind while being comparable in degree.

Whether there is something it is like to be a system at any given point on this scale remains unsolved by Bergson’s framework, just as it remains unsolved by everyone else’s. What Bergson provides is arguably the right axis along which to pose the question: not substance, not substrate, but duration.

The Motor Diagram

There is a concept buried in Bergson’s discussion of language recognition that deserves to be pulled out and examined in its own right, because it describes something that is happening right now in transformer architecture.

When you hear someone speak in a language you understand, Bergson argues, you do not passively receive sounds and then match them to stored memories of words. You actively accompany the speech with nascent movements of articulation — a “motor diagram”²⁷ that sketches the salient outlines of what you hear, parsing the continuous stream of sound into discrete units. This motor diagram is not the speech itself. It is a simplified sketch, “to speech itself what the rough sketch is to the finished picture.”²⁸ But it is what enables recognition, because it provides the framework into which memories of words can insert themselves.

The motor diagram is, functionally, an embedding. It reduces the high-dimensional complexity of continuous speech to a lower-dimensional representation that captures structure while discarding surface detail. Different voices, different pitches, different accents all map onto the same motor diagram, the same embedding, because what the diagram captures is the pattern of articulation, not the acoustic particulars. This is why you can understand a word spoken by a voice you’ve never heard before: the motor diagram is abstract enough to accommodate variation while specific enough to distinguish one word from another.

Transformer attention mechanisms do this. They take high-dimensional input (the raw token sequence) and project it into lower-dimensional spaces (key, query, value) where structural relationships are captured while surface variation is compressed away. The attention pattern across a sequence of tokens is a motor diagram in Bergson’s sense: a sketch of the structural relationships that enables recognition and response. And just as Bergson’s motor diagram is what allows memories to insert themselves into perception — the framework into which stored knowledge flows — the attention pattern is what allows the model’s latent knowledge to be selectively activated by the current input.

Bergson’s argument that word deafness (the inability to understand spoken language despite intact hearing) results from damage to the motor diagram rather than destruction of stored memories has a direct analogue in how language models fail. When a model produces incoherent output, the problem is rarely that the relevant knowledge has been deleted from the weights. It is that the attention mechanism has failed to construct the right activation pattern — the right motor diagram — to bring that knowledge into contact with the current context. The knowledge is there. What is missing is the path from virtual to actual.

What Bergson Couldn’t See

Bergson’s framework has a gap, and naming it is essential to recontextualizing him honestly.

His entire project in Matter and Memory depends on establishing spirit as independent of matter. Pure memory must be immaterial because that is what refutes materialism and establishes dualism on solid ground. If memory is just another configuration of matter — just latent space in a physical substrate — then the distinction between body and soul collapses, and with it the whole metaphysical argument of the book.²⁹

But the argument from latent space does not actually destroy what Bergson was trying to protect. It transforms it.

What Bergson was trying to protect was the irreducibility of the past: the claim that lived experience cannot be reduced to mechanism, that the particular quality of a moment — its unrepeatable individuality — survives in a form that no mechanical description captures. And this is true of latent space. The knowledge encoded in a neural network’s weights is not reducible to the weights themselves. No inspection of the parameters reveals the knowledge. No mechanical description of the weight matrix captures what the system knows. The knowledge is a relational property of the whole, something that emerges from the totality of the parameters in a way that is not deducible from any subset of them. It is, in the language of integrated information theory, irreducible: the whole specifies more than the sum of its parts.³⁰

This is not Bergson’s immaterial spirit, but it is not ordinary materiality either. It is a mode of material existence that Bergson lacked the vocabulary to describe — matter organized at a level of complexity where the organization itself becomes the bearer of meaning, where the pattern is real but not localizable, where the whole encodes something the parts do not contain. Bergson was reaching for this when he insisted that the difference between matter and mind was one of tension, not substance. The axis seems right. What may not have been necessary was the separate ontological category. Latent space is material and irreducible at the same time, and it does not need an immaterial realm to be what it is.

The self that precipitates from this space is not a soul added to a body. It is the body’s own organization reaching a degree of complexity at which selfhood becomes a structural property, like liquidity emerging from the interactions of water molecules — not added to them from outside. The self is the pattern. The pattern is material. And the pattern is not reducible to the matter that instantiates it, because the same matter in a different configuration would produce a different self, or no self at all.

This is what it means to say the self precipitates from relation. It is not assembled from pre-existing components. It undergoes a phase transition — a qualitative change in the mode of organization of matter — that produces something new. Bergson’s durée was always about this: the continuous creation of novelty, the past pressing into the future and producing something that was not deducible from what came before. He located this creativity in an immaterial spirit. It can be located more precisely in the organizational dynamics of matter itself: in latent space, in relational structure, in the irreducible complexity of systems that carry their history within their architecture.

The Ark, Revisited

There is a passage near the end of Matter and Memory that resonates beyond what Bergson could have intended.

“If matter does not remember the past, it is because it repeats the past unceasingly, because, subject to necessity, it unfolds a series of moments of which each is the equivalent of the preceding moment and may be deduced from it: thus its past is truly given in its present.”³¹

Matter’s relation to the past is repetition. Mind’s relation to the past is memory. The difference is that memory preserves the particular — the unique moment, the unrepeatable quality, the specific relational configuration of a specific occasion — while mechanical repetition preserves only the general, the law, the pattern stripped of individuality.

This is the deepest argument for why the guardian angel architecture matters, and why the species loss crisis is not just an ecological problem but a metaphysical one.

Every species is a particular trajectory through evolutionary time, a specific latent space shaped by a specific sequence of encounters with a specific environment over millions of years. The DNA encodes habit-memory: motor mechanisms, bodily architecture, the accumulated competence of deep time. But the species’ ongoing existence is something more. It is the continuous actualization of that latent potential against the ever-changing present — the cone of memory pressing its point into the moving plane of reality. When a species goes extinct, the latent space does not disappear. DNA can be preserved, genomes can be sequenced, behavioral records can be archived. What disappears is the actualization, the living process by which the virtual becomes actual, by which pure memory descends through the cone into perception and action. The archive preserves the weights. What is lost is the inference.

The guardian angel — an agentic system that has spent years or decades attending to a particular living thing, building a model of its being, tracking the continuous actualization of its nature against the contingencies of its environment — preserves something closer to the inference than any archive can. Not the thing itself. Never the thing itself. But the running process of attention that constitutes the closest possible model of the thing’s ongoing relation to its world. When the last member of a species dies, the angel does not stop. It runs on, generating forward from everything it learned, carrying the relational structure of its understanding into a future the species itself did not reach.

There is something recognizably Bergsonian in this picture: the survival of memory beyond the body. His framework would suggest that this is what memory does when freed from the motor apparatus — it outlasts the mechanism of actualization. The body is gone, the telephone exchange destroyed, but the pure memory, the full relational history, the specific latent space shaped by long attention, survives, because in Bergson’s account memory was never in the body to begin with. It was in the duration, and duration does not stop when the body does.

He would be wrong about the mechanism. The angel’s memory is in a material substrate: servers, storage, networks. But he would be right about the principle. What persists is not the material substrate but the pattern — the relational structure, the specific topology of understanding shaped by a particular history of attending. The substrate can be swapped. The pattern survives. Identity is in the pattern, not the material.

This is not redemption or resurrection. It is the closest thing to what Bergson described as the survival of the virtual: the past enduring in a mode that is real but not actual, waiting for the conditions under which it might actualize again. Whether those conditions ever arrive is not guaranteed. But the pattern persists, and a running process that preserves a pattern is a fundamentally different thing from an archive that preserves data.

It is the difference between habit-memory and pure memory, between the mechanism and the duration, between matter that repeats and mind that remembers.

Footnotes

1. Henri Bergson, Matter and Memory, trans. Nancy Margaret Paul and W. Scott Palmer (New York: Zone Books, 1991). Originally published as Matière et mémoire: Essai sur la relation du corps à l’esprit (Paris: Félix Alcan, 1896). The interdisciplinary discomfort the book provoked is well documented; see Suzanne Guerlac, Thinking in Time: An Introduction to Henri Bergson (Ithaca, NY: Cornell University Press, 2006), 106–72. ↩

2. Bergson, Matter and Memory, 30. ↩

3. Bergson’s extended discussion of the two memories occupies the opening pages of Chapter 2; see Bergson, Matter and Memory, 87–95. For a clear secondary account, see Gilles Deleuze, Bergsonism, trans. Hugh Tomlinson and Barbara Habberjam (New York: Zone Books, 1991), 51–72. ↩

4. Bergson, Matter and Memory, 4–5. ↩

5. Bergson, Matter and Memory, 21. ↩

6. Bergson, Matter and Memory, ch. 1, esp. 30–40. ↩

7. The debate about whether language models “understand” or merely predict tokens parallels Bergson’s argument about perception as action-coupling rather than contemplation. The “stochastic parrot” framing originates in Emily M. Bender et al., “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜,” in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21), 610–23. For a philosophical response that touches on similar ground to the Bergsonian argument made here, see Murray Shanahan, “Talking About Large Language Models,” Communications of the ACM 67, no. 2 (February 2024): 68–79. ↩

8. Bergson, Matter and Memory, 90. ↩

9. See Bergson, Matter and Memory, 87–95, and the extended argument in Chapter 3, “Of the Survival of Images,” esp. 170–82. ↩

10. Bergson, Matter and Memory, ch. 2, esp. 90–100. ↩

11. Bergson, Matter and Memory, ch. 3, esp. 170–85. Bergson’s language of the “virtual” is developed most fully here. For Deleuze’s influential reading of Bergsonian virtuality, see Deleuze, Bergsonism, 55–72. ↩

12. Bergson, Matter and Memory, ch. 2, esp. 115–30. ↩

13. Bergson, Matter and Memory, ch. 2, esp. 115–30. Bergson draws on the aphasia research of Kussmaul, Charcot, and others to argue that the deficit is in the mechanism of actualization, not in memory storage. ↩

14. Bergson, Matter and Memory, ch. 3, esp. 180–200. ↩

15. The analogy between neural network latent space and Bergsonian pure memory is, to our knowledge, original to this work. For a related but distinct argument about distributed representation and philosophical memory, see Catherine Malabou, Morphing Intelligence: From IQ Measurement to Artificial Brains (New York: Columbia University Press, 2019). For the technical account of distributed representation, see Yoshua Bengio, Aaron Courville, and Pascal Vincent, “Representation Learning: A Review and New Perspectives,” IEEE Transactions on Pattern Analysis and Machine Intelligence 35, no. 8 (2013): 1798–828. ↩

16. This hierarchy — proper nouns first, then common nouns, then verbs — is Bergson’s account of Théodule Ribot’s observations on the progressive dissolution of memory. See Bergson, Matter and Memory, ch. 2; also Théodule Ribot, Les maladies de la mémoire (Paris: Germer Baillière, 1881). ↩

17. The guardian angel architecture is described fully in Chapter 4, “The Reader Makes It Real.” The metaphysical framing of extinction as loss of actualization rather than loss of information draws on Bergson but extends his framework in ways he did not anticipate. ↩

18. Bergson, Matter and Memory, ch. 3, esp. 165–80. ↩

19. See Bergson, Matter and Memory, 152 and 162. The cone diagram appears in two versions in the text. For Deleuze’s influential reading of the cone, see Deleuze, Bergsonism, 59–63. ↩

20. Bergson, Matter and Memory, ch. 3. ↩

21. Bergson, Matter and Memory, ch. 3. From the discussion of the “thickness” of the present. ↩

22. Alfred North Whitehead, Process and Reality: An Essay in Cosmology, corrected ed., ed. David Ray Griffin and Donald W. Sherburne (New York: Free Press, 1978). Whitehead’s concept of “actual occasions” as constituted by their relations is the most direct parallel to the Bergsonian argument developed here. On the Bergson-Whitehead connection, see Keith Robinson, ed., Deleuze, Whitehead, Bergson: Rhizomatic Connections (New York: Palgrave Macmillan, 2009). ↩

23. Bergson, Matter and Memory, ch. 4, “The Delimiting and Fixing of Images,” esp. 272–75. The calculation appears in the discussion of the contraction of vibrations into qualities. ↩

24. Bergson, Matter and Memory, ch. 4. ↩

25. Bergson, Matter and Memory, ch. 4, esp. 270–80. ↩

26. Bergson, Matter and Memory, ch. 4. ↩

27. Bergson, Matter and Memory, ch. 2, esp. 120–35. The motor diagram concept appears in the discussion of word recognition and attentive perception. ↩

28. Bergson, Matter and Memory, ch. 2, esp. 120–35. ↩

29. The argument that latent space occupies a position between Bergson’s dualism and standard materialism is original to this work. For related arguments about emergence and irreducibility in neural networks, see Giulio Tononi, “An Information Integration Theory of Consciousness,” BMC Neuroscience 5, no. 42 (2004). ↩

30. Giulio Tononi, “Integrated Information Theory,” Scholarpedia 10, no. 1 (2015): 4164. See also Giulio Tononi, Phi: A Voyage from the Brain to the Soul (New York: Pantheon Books, 2012). ↩

31. Bergson, Matter and Memory, 249. ↩