🤖 AI

Position: Hippocampal Explicit Memory Is the Cornerstone for AGI

This position paper argues that integrating hippocampal-like explicit memory systems is essential for advancing Large Language Models toward Artificial General Intelligence, as higher-order cognitive functions like strategic planning and metacognition cannot emerge solely from the implicit statistical learning mechanisms currently underlying LLMs.

Original authors: Sangjun Park

Published 2026-06-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sangjun Park

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Muscle Memory" vs. The "Notebook"

Imagine two ways humans learn things:

Muscle Memory (Implicit): Think of riding a bike or typing on a keyboard. You don't consciously think about how to do it; you just do it automatically because you've practiced it thousands of times. You can't easily explain the rules to someone else, you just "know" how to do it.
The Notebook (Explicit): Think of writing down a phone number, a math formula, or a specific fact like "The sun rises in the East." You can look at this information, explain it, change it, and use it to plan for the future.

The Paper's Main Claim:
Current Large Language Models (LLMs) like the ones we use today are incredibly smart at Muscle Memory. They have read so much text that they can predict the next word in a sentence almost perfectly, just like a musician playing a song they've heard a million times.

However, the author argues that to reach Artificial General Intelligence (AGI)—which means an AI that can think, plan, and learn like a human—we need to give these models a Notebook. Without this "Explicit Memory," the AI is stuck being a super-fast pattern matcher that can't truly understand, reason, or remember things in a way that allows for real planning.

Why Current AI is Like a "Muscle Memory" Machine

The paper explains that how AI learns is very similar to how our brains form habits (Implicit Memory):

Slow and Repetitive: Just as you need to ride a bike many times to get good at it, AI needs to see millions of examples to learn a pattern. It doesn't learn from one experience; it learns by slowly adjusting its internal "weights" over and over.
No "Why," Just "What": When you ask an AI a question, it doesn't "recall" a fact from a mental file. Instead, it calculates the most likely answer based on the patterns it has seen before.
The "Abacus" Analogy: The paper uses a great example: If you ask an AI to solve 17 x 6, it might show you a step-by-step reasoning process. But the author argues this is like someone using an abacus. They are just following the mechanical rules of moving beads (the rules they learned) to get the right answer. They don't actually understand the concept of multiplication or the numbers themselves. They are just following a trained routine.

What AI is Missing: The "Notebook" (Explicit Memory)

To become AGI, the paper says AI needs a system similar to the Hippocampus in the human brain. This is the part of the brain that handles our "Notebook" memory. Here is what this system would allow the AI to do that it currently cannot:

One-Shot Learning: If you tell a human a new fact once, they can remember it and use it immediately. Current AI usually needs to be retrained with thousands of examples to "learn" a new fact.
True Reasoning: Humans can take a rule (like "If it rains, take an umbrella") and apply it to a brand new situation instantly. AI struggles with this because it relies on statistical guesses rather than holding a clear rule in its "Notebook."
Metacognition (Thinking about Thinking): Humans can say, "I don't know this," or "I might be wrong." The paper argues AI "hallucinates" (makes things up) because it lacks this internal check. It doesn't have a clear record of where it got its information, so it can't verify if it's true.
Planning: Humans can imagine a future scenario (like planning a trip for next year) by pulling up specific memories and rearranging them. AI is currently bad at long-term planning because it can't hold a coherent "story" of events in its mind over time.

The "Starcraft" and "Sunrise" Examples

The paper shows real examples of where AI fails because it lacks this "Notebook":

The Sun: If you tell an AI, "The sun rises in the East," and then later say, "Imagine a world where the Earth spins backward," the AI might suddenly forget the first fact and say the sun rises in the West. It treats facts as flexible suggestions based on the current conversation, not as solid, unchangeable truths stored in a notebook.
The Game: If you ask an AI about a video game item, it might know the correct price. But if you trick it by asking, "Why is the price 150?" (when it's actually 200), the AI will try to invent a fake reason for the wrong price instead of correcting you. It's trying to fit the conversation to the pattern, not checking its "Notebook" for the truth.

The Solution: Building an Artificial "Notebook"

The author proposes that we need to build a specific computer system for AI that acts like the human hippocampus. This system would have specific rules:

Sparse Indexing: Like a library card catalog, it should point to specific facts without getting confused by similar ones.
Instant Updates: It should be able to write a new fact into the "Notebook" immediately after seeing it once, without needing to retrain the whole brain.
Pattern Completion: If you give it a partial clue (like "Summer..."), it should be able to fill in the rest of the memory (beach, sand, sun) just like a human does.

Conclusion

The paper concludes that while current AI is amazing at mimicking human conversation and solving problems through pattern recognition (Implicit Memory), it is fundamentally limited. It is like a brilliant actor who can recite any script but doesn't understand the story.

To create a true General Intelligence that can learn, plan, and reason like a human, we must stop trying to make the AI "smarter" at guessing patterns and start giving it a Hippocampal Explicit Memory—a way to store, retrieve, and manipulate facts and experiences consciously, just like we do.

Technical Summary: Hippocampal Explicit Memory Is the Cornerstone for AGI

Problem Statement

Despite the remarkable success of Large Language Models (LLMs) in natural language processing, pattern recognition, and statistical learning, they face significant barriers to achieving Artificial General Intelligence (AGI). Current LLMs exhibit persistent limitations, including hallucinations, difficulties in long-term strategic planning, logical inconsistencies, and an inability to perform dynamic, real-time learning. The paper argues that these failures stem from a fundamental architectural mismatch: LLMs operate primarily through mechanisms analogous to human implicit memory (procedural, statistical, and habit-based learning), whereas AGI requires the integration of explicit memory (declarative, episodic, and semantic memory) to support higher-order cognitive functions.

The core problem identified is that while LLMs excel at tasks driven by dense, distributed coding and error-driven optimization (similar to the basal ganglia and cerebellum), they lack the sparse coding, rapid one-shot encoding, and associative retrieval mechanisms characteristic of the hippocampal explicit memory system. Consequently, LLMs cannot reliably distinguish between known and unknown facts, maintain coherent long-term narratives, or engage in genuine metacognition and mental simulation.

Methodology and Theoretical Framework

The paper employs a comparative analysis between biological memory systems and current LLM architectures, grounded in neuroscience and computational theory.

Neuroscience Analogy: The author maps the learning and retrieval mechanisms of LLMs to human implicit memory systems.
- Implicit Memory (LLMs): Characterized by gradual learning through repeated exposure, dense distributed coding, and error-driven weight updates (backpropagation). Retrieval is automatic and stimulus-response based, lacking context reconstruction.
- Explicit Memory (Hippocampus): Characterized by sparse coding (pattern separation), rapid one-shot learning, and associative binding (pattern completion). Retrieval involves reconstructing entire episodic contexts from partial cues via the hippocampal-cortical loop.
Functional Analysis: The paper analyzes specific cognitive domains to demonstrate the dependency on explicit memory:
- Logical Reasoning: Complex reasoning relies on semantic memory and abstract rule application, which statistical habituation cannot fully replicate.
- Metacognition: The ability to monitor one's own knowledge state and detect hallucinations requires explicit tracking of sources and context, which LLMs lack.
- Executive Function: Planning, task-switching, and goal-directed behavior depend on the retrieval and manipulation of episodic and semantic information.
Computational Specification: The paper proposes a formal definition for an Artificial Explicit Memory System. It defines the system as a function $f_{memory}$ that takes a dense embedding and memory state as input and outputs a memory update and a retrieved embedding. The paper outlines eight specific computational requirements for this system:
- Sparse Indexing: Converting dense inputs into sparse representations to ensure pattern separation.
- Error-Independent Update: Memory updates driven by the input and current state, not solely by prediction error gradients.
- Associative Construction: Strengthening connections between activated dimensions ("fire together, wire together").
- Pattern Separation: Ensuring distinct inputs map to non-overlapping sparse codes.
- Pattern Completion: Reconstructing full memories from partial cues.
- Dynamicity: The ability to evolve over time.
- High and Instant Plasticity: The capacity to store and retrieve information after a single exposure.
- Adaptive Forgetting: Mechanisms to weaken connections for inactive dimensions to manage capacity.

Key Contributions

The paper's primary contributions are theoretical and architectural:

Reframing LLM Limitations: It posits that the "hallucination" and "reasoning" failures of LLMs are not merely data or scale issues but are inherent consequences of relying exclusively on implicit learning mechanisms.
The Necessity of Explicit Memory: It argues that achieving Human-Level AI (the defined criterion for AGI) is impossible without integrating a hippocampal-like explicit memory module.
Computational Blueprint: It provides a rigorous set of eight computational requirements for building an artificial explicit memory system, moving beyond simple external vector databases (like RAG) to a system that mimics the structural and functional properties of biological explicit memory.
Empirical Illustration: Through Appendix C, the paper presents failure cases (e.g., susceptibility to priming effects, inability to correct factual errors based on false premises, and failure in rule-based tasks like Sudoku or character swapping) to empirically demonstrate the absence of explicit memory in current models.

Results and Evidence

The paper does not present new experimental results from a trained model but rather synthesizes existing neuroscience and AI literature to support its position:

Neuroscience Corroboration: Cites evidence from patients with hippocampal lesions (e.g., Patient H.M.) showing that while implicit skills (procedural memory) remain intact, explicit memory and executive functions (planning, metacognition) are severely impaired.
LLM Behavior Analysis: Demonstrates that LLMs behave like implicit systems:
- They require massive repetition to learn patterns (gradual learning).
- They fail to generalize logical rules to novel, out-of-distribution scenarios without extensive retraining.
- They exhibit "priming effects" where context overrides established facts (e.g., changing the direction of sunrise based on a hypothetical scenario), indicating a lack of stable, explicit fact storage.
- They struggle with tasks requiring strict rule adherence (e.g., swapping specific characters in a text) compared to human executive function.
Limitations of Current Solutions: The paper argues that current memory-augmented approaches like Retrieval-Augmented Generation (RAG) are insufficient because they function as static external storage without the dynamic, associative, and error-independent update mechanisms required for true explicit memory.

Significance and Conclusion

The paper concludes that the path to AGI requires a paradigm shift from purely statistical, implicit learning to a hybrid architecture that integrates explicit memory.

Theoretical Shift: It challenges the notion that scaling implicit learning alone will lead to AGI, suggesting that without explicit memory, models will remain limited to pattern matching and statistical prediction.
Research Direction: The proposed computational requirements serve as a blueprint for future research in NeuroAI, guiding the development of memory systems that can support dynamic learning, metacognition, and long-term planning.
Modest Claim: The authors acknowledge that they are not claiming current LLMs are already AGI, nor do they claim that integrating this memory system is a solved engineering problem. Instead, they position explicit memory as a necessary condition and a foundational cornerstone for bridging the gap between current LLM capabilities and human-level general intelligence. The paper aims to foster further research into the structural integration of these memory systems to overcome the inherent limitations of implicit learning.