arXiv:2606.11245 Claims Hippocampal Explicit Memory Required to Surpass LLM Scaling Limits

The June 2026 position paper from Sangjun Park states that LLM training parallels human implicit memory while AGI-level functions depend on hippocampal explicit memory for binding episodes, enabling retrieval, and supporting offline consolidation. It outlines computational requirements including structured storage, content-addressable access, and separation of fast and slow learning systems. Current agent frameworks such as ReAct, Reflexion, and Voyager rely on prompt buffers or vector stores that lack the binding and schema formation properties of hippocampal circuits. These additions produce measurable gains on short tasks yet fail to close the gap on multi-day planning benchmarks where performance saturates despite increased model scale. Neuroscience evidence shows hippocampal lesions impair relational inference and prospective simulation while leaving perceptual pattern recognition intact, mirroring the dissociation between LLM capabilities and AGI requirements. Architectural analyses of transformer scaling laws indicate that additional parameters improve interpolation within training distributions but do not add the variable-binding mechanisms needed for out-of-distribution strategy generation. Next steps require hybrid systems that maintain an explicit episodic store updated via fast weights, with consolidation rules tested against neuroscience-derived metrics rather than task accuracy alone. Without such separation, continued scaling will encounter hard limits on long-horizon coherence.