ICML paper argues explicit memory is LLMs' missing link to AGI

Sangjun Park's ICML 2026 position paper makes a bold neuroscientific claim: today's LLMs are stuck in a cognitive cul-de-sac because they operate solely on implicit memory. Drawing direct parallels between the statistical learning of transformers and the brain's unconscious pattern recognition, Park argues that explicit memory—modeled after the hippocampus—is the necessary component for AGI. Without it, LLMs cannot perform long-term strategic planning, metacognition, or abstract symbolic reasoning, all of which are hallmarks of human higher-order cognition. The paper is not just theoretical; it proposes concrete computational frameworks for designing artificial explicit memory systems that could be integrated into existing architectures.

Park's argument arrives at a time when many AI labs are hitting diminishing returns from scaling parameters alone. The paper suggests that the next leap forward may require a fundamentally different architecture—one that separates learned patterns (implicit) from episodic, recallable events (explicit). This hybrid approach could unlock agents that remember past decisions, reason over long timelines, and self-correct. While the paper is a position rather than an implementation, it provides a clear research direction for the field, bridging neuroscience and AI engineering.