Overview
A position paper accepted to the ICML 2026 Position Paper Track argues that integrating explicit memory—modeled after the human hippocampus—is the essential missing component for advancing Large Language Models (LLMs) toward Artificial General Intelligence (AGI).
Key Argument
The paper contends that while LLMs have achieved impressive performance across numerous tasks, their underlying learning mechanism is fundamentally analogous to human implicit memory. Implicit memory supports pattern recognition and statistical learning but lacks the capacity for higher-order cognitive functions required for AGI. These include long-term strategic planning, metacognition, and symbolic reasoning—all of which depend heavily on the hippocampal explicit memory system in the human brain.
Neuroscientific Foundation
Drawing on findings from neuroscience, the author establishes that:
- LLMs learn implicitly: They capture statistical regularities in data, similar to how humans acquire procedural skills and priming.
- Explicit memory is distinct: The hippocampus enables conscious recollection of facts and events, supporting flexible reasoning and future planning.
- Cognitive functions require both systems: AGI demands not just pattern recognition but also deliberate, context-aware reasoning that explicit memory enables.
Computational Implications
The paper bridges neuroscience and AI by outlining computational requirements for artificial explicit memory systems:
- Episodic storage: Maintain distinct, retrievable records of past experiences.
- Rapid encoding and consolidation: Efficiently transfer new information from short-term to long-term storage.
- Flexible retrieval: Access relevant memories based on context, not just statistical similarity.
- Integration with implicit systems: Allow seamless cooperation between learned patterns and explicit recollections.
2026 Context
As of 2026, the field is actively exploring hybrid architectures that combine large-scale neural networks with memory-augmented components. This paper adds a critical neuroscientific perspective, advocating for explicit memory as the foundational building block—rather than just an optional enhancement—for achieving AGI.
Conclusion
The author hopes this perspective will foster further interdisciplinary research and lay the groundwork for integrating explicit memory into next-generation AI systems, bringing them closer to human-like general intelligence.
via ArXiv AI