Big-Loop Recurrence Within the Hippocampal System Supports Integration of Information Across Episodes

Type: paper Slug: big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes—hassabis Sources: big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes—hassabis Last updated: 2026-05-13

---

Summary

Koster, Chadwick, Chen, Berron, Banino, Düzel, Hassabis, and Kumaran (2019) used fMRI pattern similarity analysis to demonstrate that the hippocampal system integrates information across separately experienced episodes via “big-loop” recurrence — connections between hippocampus and neocortical areas. When participants encountered objects that had appeared in different contexts, hippocampal-neocortical pattern similarity increased, suggesting a mechanism for how disjoint experiences are linked into coherent knowledge.

Core content

Problem: How does the brain integrate information from separate experiences into unified knowledge? Standard hippocampal models focus on within-episode binding, but real-world knowledge requires cross-episode integration (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Introduction).

“Big-loop” vs. “small-loop” recurrence:

• Small-loop: Hippocampal subfield interactions (CA3→CA1→subiculum) — supports within-episode binding (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Introduction).
• Big-loop: Hippocampus↔neocortex interactions (entorhinal cortex, parahippocampal cortex, retrosplenial cortex) — proposed to support cross-episode integration (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Introduction).

Experimental design: Participants encoded objects in different spatial contexts across separate episodes. fMRI measured pattern similarity in hippocampal and neocortical regions when the same object appeared in new combinations, testing whether the brain linked information across episodes (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Methods).

Key findings:

• Pattern similarity in hippocampal-neocortical (big-loop) regions increased for objects that shared a context, even when encountered in separate episodes (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Results).
• This effect was specific to neocortical targets of hippocampal projections (entorhinal, parahippocampal) and not found in unrelated cortical regions (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Results).
• The degree of cross-episode pattern similarity predicted behavioral integration performance (paper—big-loop-recurrence-within-the-hippocampal-system-supports-integration-of-information-across-episodes §Results).

Connections- Theme: theme—hippocampal-construction

• Project: hippocampus-research
• Collaborators: Raphael Koster (co-first), Martin J. Chadwick (co-first), Yi Chen, David Berron, Andrea Banino, Emrah Düzel, Dharshan Kumaran
• Era: alphafold-era
• Venue: venue—Neuron
• Related: paper—the-construction-system-of-the-brain — hippocampus as construction system, this provides mechanistic detail
• Related: paper—what-learning-systems-do-intelligent-agents-need-complementary-learning-systems-theory-updated — cross-episode integration as a CLS mechanism

Honest Gaps

• Metadata lists Kumaran, Vigliocco, Burgess as co-authors; actual authors are Koster, Chadwick, Chen, Berron, Banino, Düzel, Hassabis, Kumaran. None of Vigliocco or Burgess are authors; Koster, Chadwick, Chen, Berron, Banino, Düzel are all missing. This is a complete metadata error.
• The fMRI evidence shows correlational pattern similarity — it cannot establish causal direction (does hippocampus drive neocortical integration, or vice versa?).
• The “big-loop” terminology, while evocative, is an anatomical simplification — the actual circuitry involves multiple parallel pathways.
• Sample size was modest (typical for fMRI of this era).
• The task uses artificial object-context pairings — generalization to naturalistic cross-episode integration is unclear.