Encoding, transduction, and world models — Part 3/3

Part 3/3 of the article “Encoding, transduction, and world models”. Follows part 1/3 (chapters 1-2) and part 2/3 (chapters 3-5). Original article in French.

6. Memory as polycentric graph: towards a biographical architecture of cognition

Human memory is not a unitary system but a polycentric graph — a network of distributed engrams (Tonegawa et al., 2015) where each node is simultaneously a destination and a starting point. The edges of this graph do not reflect statistical co-occurrence but lived co-experience, modulated by affective valence (Damasio’s somatic markers) and temporally indexed. Synaesthetic memory constitutes a six-sense weaving of the fabric of existence: the distributed engram is not a stored entity, it is a texture navigable from multiple perspectives (Nigro & Neisser, 1983). The biographical edge — a relation between two representations encoding their co-activation within a single episode lived by a continuous agent — constitutes the fundamental architectural gap with current AI systems. An LLM does not weave, it indexes.

7. World models: decisive advance or still insufficient threshold?

JEPA architectures (LeCun, 2022) and embodied robotic systems (DreamerV3, RT-X) reduce epistemic distance to the world by introducing genuine sensorimotor grounding. They address sensory grounding and world modeling, but do not yet produce an integrated cognitive biography. No current architecture demonstrates the simultaneous co-presence of: persistent agent continuity, autonoetic episodic memory, affective modulation of mnesic relations, and multi-perspective navigability of mnesic space.

8–9. Epistemological implications and conclusion

The gap between human cognition and contemporary architectures is neither quantitative nor ontological — it is architectural and historical. LLMs model human representations of the world, not the world itself. Embodiment is necessary but not sufficient. The central question: can we conceive a system whose intelligence would be founded not only on data but on a history of interactions integrated into an autobiographical memory — a system that has something to remember because it lived it?

Read the document

Access the full article

Enter your details to access the document. Free access — no sales outreach.

Personalized document · Free access · No sales outreach

Key takeaways

Human memory is a polycentric graph: a network of distributed engrams with no root, where each node is simultaneously a destination and a potential starting point depending on activation context.
The biographical edge encodes the co-activation of two representations within a single episode lived by a continuous agent — with temporal indexing, affective modulation, and contextual retrieval.
An LLM does not weave, it indexes. Its edges are statistical co-occurrence edges; those of human memory are lived co-experience edges.
Synaesthetic memory is a six-sense weaving of the fabric of existence. The distributed engram is not a stored entity: it is a texture.
World models (JEPA, LeCun 2022) reduce epistemic distance to the world but only solve sensory grounding and world modeling — not biographical memory.
No current architecture demonstrates the simultaneous co-presence of: agent continuity, autonoetic episodic memory, affective modulation of links, multi-perspective navigability.
The gap between human cognition and AI systems is neither quantitative nor ontological — it is architectural and historical. The question: can we build a system that has something to remember because it lived it?