Body Memory and the Limits of Statistical Learning: Thomas Fuchs on What AI Cannot Train On

Framing the Problem

The May 2026 deep analysis established that Merleau-Ponty's motor intentionality and current predictive AI architectures are structurally homologous — both are forward models — but differ in three fundamental respects: the content of their predictions (felt vs. visual), the history through which they were built (embodied experience vs. population statistics), and the evaluative dimension of their operation (affordance sensitivity vs. objective minimisation).

The second of these differences — the grounding in embodied experience vs. population statistics — deserves more precise analysis than a single comparative observation allows. What exactly is embodied experience, and why can it not be reproduced by training on recordings of the body's visible output?

The phenomenologist Thomas Fuchs provides the most careful answer available in the philosophical literature. His account of body memory — developed across a series of essays and most systematically in Ecology of the Brain (2018) — identifies five distinct forms of memory that are instantiated in the body rather than in propositional cognitive content. Each form represents a different way in which lived experience sediments in the body and shapes future movement. Taken together, they define what an AI system trained on video of human movement cannot access — and what sensing modalities might, in principle, approach.

The Five Forms of Body Memory

Fuchs distinguishes five modes of body memory, each corresponding to a different kind of embodied learning:

1. Procedural memory — the body's retention of learned motor skills as directly executable capacities. Playing an instrument, riding a bicycle, executing a dance phrase: these are not stored as propositional knowledge ("first move the bow in direction X at velocity Y") but as bodily habits that can be enacted without deliberation. The knowledge is in the movement, not in a mental representation of the movement. Merleau-Ponty's motor intentionality is the active expression of procedural memory.

2. Situational memory — the body's ability to recognise and respond to familiar situations before conscious identification occurs. The familiar smell of a rehearsal studio, the acoustic quality of a performance space, the temperature and humidity of the air: these situational cues activate appropriate movement dispositions and attentional orientations before the mover consciously registers the situation. Situational memory is why experienced performers can calibrate their movement to a new venue faster than their conscious perception could account for.

3. Intercorporeal memory — the body's retention of patterns of interaction with other bodies. This includes the somatic attunement that develops between regular movement partners: the Contact Improvisation practitioners who have worked together for years and who can anticipate each other's momentum before it becomes visible, the teacher and student whose physical interaction has developed a mutually calibrated sensitivity. Intercorporeal memory is not stored in one body — it exists between bodies, in the patterns of interaction that two movement histories have negotiated.

4. Incorporative memory — the body's retention of external objects and tools as extensions of the body schema. Merleau-Ponty's example of the stick that becomes an extension of the blind person's sense of touch is a paradigm case: through practice, the tool is incorporated into the body schema, and its distal end becomes the locus of tactile attention. For musicians, the instrument; for dancers working with props or partners; for practitioners using wearable sensors or interfaces: the incorporative extension of the body schema is a form of memory that persists across sessions of use.

5. Traumatic or pain memory — the body's retention of injury, shock, and chronic pain in patterns of movement avoidance, protective muscular bracing, and altered kinaesthetic sensibility. This is body memory in its most constraining form: the body that has been injured reorganises its movement vocabulary around the protection of injured areas, often in ways that outlast the original injury and become habitual adaptations. For somatic practitioners, trauma memory is one of the primary objects of therapeutic work — the unwinding of protective patterns that have become unnecessary but persist.

What Body Memory Is Not

Fuchs is precise about what distinguishes these five forms of body memory from other memory systems. They are not:

  • Episodic memory (the explicit recollection of past events as events). Body memory does not present its content as a recalled episode; it is enacted directly in current movement without accompanying awareness of the past.

  • Semantic memory (stored propositional knowledge). Body memory is not "knowing that" something is the case; it is "knowing how" to do something, where the knowledge is inseparable from the doing.

  • Representation in the cognitive-science sense. Body memory does not encode a representation of past movement that the body then decodes and executes. It is a modification of the body's current capacities — a change in what the body can do, not a change in what it knows about what it once did.

This is the philosophically critical point. Standard AI architectures are representation systems: they encode patterns from training data and use those patterns to generate outputs. Even the most sophisticated current models are representation-based. Body memory, in Fuchs's account, is not representational: it is constitutive. It does not represent past movements; it is the past movements, sedimented in the body's current capacities.

The Training Data Problem, Precisely Stated

With Fuchs's five forms in view, the limitation of training AI motion systems on video data can be stated precisely rather than gesturally.

Video data captures the visible output of movement. What it does not capture:

  • Procedural memory — the felt organisation of the movement, the proprioceptive and kinaesthetic experience that constitutes the skill. The video records what the skill produced, not the skill itself.

  • Situational memory — the contextual cues (acoustic, haptic, olfactory, thermal) that activated the movement response. The video records the response; the situational trigger is invisible to the camera.

  • Intercorporeal memory — the history of partner interaction that makes a specific Contact Improvisation partnership qualitatively different from any other. The video records what two bodies do together; it cannot record the accumulated relational history that produced that specific quality of attunement.

  • Incorporative memory — the extended body schema that incorporates tools, partners, or interfaces as felt extensions of the self. The video records the mover with their tool or partner; it cannot record the incorporation of that tool into the mover's body schema.

  • Traumatic memory — the protective and adaptive patterns that are the body's retained response to past injury. These patterns appear in movement data as systematic deviations from "normal" biomechanics, but the video cannot reveal that they are responses to past trauma rather than simply idiosyncratic movement preferences.

A model trained on video is trained on the visible outputs of these five forms of body memory, without access to the memory processes themselves. This is not a limitation of training data size or quality. It is a structural limitation: the data modality does not encode what is most significant about human movement from the perspective of somatic intelligence.

What Alternative Sensing Modalities Could Offer

If the problem is that video data captures outputs while body memory is constituted in processes, then the research question is: what sensing modalities are closer to the processes?

EMG (surface electromyography) captures the electrical activity of muscles — the neuromuscular production process that generates movement before it is visible. This is closer to procedural memory: it records something of the skill's neuromotor organisation rather than only its visible outcome. An AI trained on EMG data from skilled and unskilled movers performing the same task would have access to something of the qualitative difference between skilled and unskilled execution — not just the positional difference.

IMU (inertial measurement unit) arrays capture accelerations and orientations — the dynamic qualities of movement (sudden vs. sustained, bound vs. free in the acceleration domain) that correspond to some of Laban's Effort qualities. This is closer to the kinetic texture of movement than video provides.

Multi-person synchronised sensing (dual EMG, dual IMU, force plates) could, in principle, begin to approach the intercorporeal dimension: recording the mutual physical accommodation between two movers' nervous systems rather than only the positions of two bodies in space. This is technically challenging and has not been attempted at the scale relevant to training motion generation models.

What none of these sensing modalities can capture — and this is the hard residue of Fuchs's analysis — is the phenomenological dimension: what the movement feels like from the inside, the first-person quality of kinaesthetic experience. This is not a gap that can be closed by adding more sensors. It is the limit of third-person scientific measurement when applied to first-person experience.

The Implication for Somatic AI Research

Fuchs's analysis does not lead to pessimism about AI motion generation. It leads to precision about what is achievable and what is not.

What is achievable: AI systems that are trained on richer, more proximal sensing data (EMG, IMU, force, multi-person synchronised) will be able to generate movement that is qualitatively more faithful to the somatic experience of the training practitioners than systems trained on video alone. The gap between generated movement and felt movement can be narrowed, even if it cannot be closed.

What is not achievable by current methods: AI systems that have body memory in Fuchs's sense — that have the accumulated first-person kinaesthetic history of a practitioner's years of embodied training. A model trained on EMG data from expert movers has learned statistical patterns in expert neuromuscular organisation. It has not undergone the practitioner's embodied development. It does not have the procedural, situational, intercorporeal, incorporative, or traumatic memory that would make its generative proposals fully commensurate with the practitioner's own.

The practical implication for somatic AI co-creation is the same one that Amy LaViers identified from the engineering direction: AI is most useful as a collaborator and amplifier for a practitioner who already has body memory, not as a substitute for the embodied development that produces body memory. The AI's generative proposals are most valuable when they are received by a body that has the somatic intelligence to respond to them — to inhabit, modify, resist, and extend them from the inside.

This is not a limitation unique to somatic AI. It is the appropriate division of labour between systems that learn from data and practitioners who learn through their bodies.

APA References

Fuchs, T. (2018). Ecology of the brain: The phenomenology and biology of the embodied mind. Oxford University Press. https://doi.org/10.1093/med/9780199646883.001.0001

Fuchs, T. (2012). The phenomenology of body memory. In S. Koch, T. Fuchs, M. Summa, & C. Müller (Eds.), Body memory, metaphor, and movement (pp. 9–22). John Benjamins.

Merleau-Ponty, M. (1962). Phenomenology of perception (C. Smith, Trans.). Routledge. (Original work published 1945)

Sheets-Johnstone, M. (2011). The primacy of movement (2nd ed.). John Benjamins. https://doi.org/10.1075/aicr.82