Hippocampal Construction

Type: theme Slug: theme—hippocampal-construction Sources: deconstructing-episodic-memory-with-construction—hassabis, patients-with-hippocampal-amnesia-cannot-imagine-new-experiences—hassabis, the-construction-system-of-the-brain—hassabis, using-imagination-to-understand-the-neural-basis-of-episodic-memory—hassabis, imagine-all-the-people-how-the-brain-creates-and-uses-personality-models—hassabis, decoding-neuronal-ensembles-in-the-human-hippocampus—hassabis, decoding-individual-episodic-memory-traces-in-the-human-hippocampus—hassabis, tracking-the-emergence-of-conceptual-knowledge-during-human-decision-making—hassabis, semantic-representations-in-the-temporal-pole-predict-false-memories—hassabis, computations-underlying-social-hierarchy-learning—hassabis, neural-mechanisms-of-hierarchical-planning-in-a-virtual-subway-network—hassabis, big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes—hassabis Last updated: 2026-05-13

---

Summary

The hippocampal construction hypothesis is the central intellectual contribution of Hassabis’s neuroscience career. It proposes that the hippocampus does not store memories as holistic recordings but instead constructs episodic representations by recombining stored elements (people, places, objects, actions) into scene-like wholes. This same constructive mechanism is claimed to support imagination, future thinking, conceptual abstraction, and planning. The hypothesis originates in the 2007 BBS target article and recurs, explicitly or implicitly, across 12 corpus publications spanning 2007–2019.

Core content

The core claim: Episodic memory is not retrieval but construction. When you remember a past event, the hippocampus assembles a scene from distributed cortical stores — it does not play back a recording. This explains why memories are reconstructive (and error-prone) and why the same system can be used for imagining events that never happened (paper—deconstructing-episodic-memory-with-construction).

Evidence layers:

1. Lesion evidence (2007): Hippocampal amnesia patients cannot imagine new experiences (paper—patients-with-hippocampal-amnesia-cannot-imagine-new-experiences) — the key dissociation.
2. Neural architecture (2008): The construction system involves hippocampus (scene construction), vmPFC (schema integration), posterior parietal cortex (spatial framework), temporal pole (semantic scaffold) (paper—the-construction-system-of-the-brain).
3. Single-neuron evidence (2009): Individual hippocampal neurons encode specific concepts that can be recombined (paper—decoding-neuronal-ensembles-in-the-human-hippocampus).
4. Ensemble-level decoding (2010): Specific episodic memories can be decoded from hippocampal ensemble patterns (paper—decoding-individual-episodic-memory-traces-in-the-human-hippocampus).
5. Behavioural paradigms (2007–2009): Imagination tasks demonstrate hippocampal involvement in constructing personality models (paper—imagine-all-the-people) and novel scenes (paper—using-imagination-to-understand-the-neural-basis-of-episodic-memory).
6. Conceptual abstraction (2009): The hippocampus tracks transitions to conceptual decision strategies (paper—tracking-the-emergence-of-conceptual-knowledge-during-human-decision-making).
7. Constructive errors (2016): Temporal pole semantic representations predict false memories — a predicted consequence of the constructive process (paper—semantic-representations-in-the-temporal-pole-predict-false-memories).
8. Hierarchical planning (2016): Hippocampal representations support hierarchical planning in a navigation task (paper—neural-mechanisms-of-hierarchical-planning-in-a-virtual-subway-network).
9. Cross-episode integration (2019): Big-loop recurrence in the hippocampal system integrates information across separate episodes (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes).

Extension to social cognition (2016): Social hierarchy learning engages hippocampal mechanisms for representing and updating self-relevant information (paper—computations-underlying-social-hierarchy-learning), extending the construction framework beyond spatial/episodic domains.

The construction-to-AI pipeline: The construction hypothesis implicitly motivates several DeepMind projects. Experience replay in DQN is constructive recombination of past transitions. GQN constructs scene representations from partial views. The neuroscience-AI bridge programme (paper—neuroscience-inspired-artificial-intelligence) makes this connection explicit.

Connections

• Theme: theme—memory-imagination, theme—episodic-memory, theme—neuroscience-AI-bridge
• Project: project—hippocampus-research
• Periods: period—phd-period (origin), period—postdoc-period (extension), period—deepmind-ascent (bridge to AI), period—alphafold-era (final papers)
• Collaborators: Eleanor A. Maguire (primary), Catriona D. Bruce, Raymond Dolan, Dharshan Kumaran, Neil Burgess
• Contrasts: theme—reinforcement-learning (model-free vs. constructive), theme—protein-folding (structurally unrelated but intellectually downstream)

Honest Gaps

• The hypothesis has never been formally tested against alternative accounts (e.g., standard consolidation theory) in a single comprehensive study.
• The neural architecture diagram (hippocampus/vmPFC/PPC/temporal pole) is based on correlational fMRI — no causal evidence for the proposed roles.
• No corpus source explicitly connects the construction system to DeepMind’s technical architecture — the link is implicit.
• The later hippocampal papers (2016, 2019) test predictions of the framework but do not substantially revise it — the theory has been static since 2008.
• No animal model work is in the corpus — all evidence is human neuroimaging or lesion studies.