# Schrodinger's Memory: Large Language Models

**Authors:** Wei Wang, Qing Li  
**arXiv:** [2409.10482](https://arxiv.org/abs/2409.10482)  
**Date:** 2024-09-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.10482v3/

## Key points

- Memory is a crucial aspect of human cognition and serves as the foundation for daily activities.
- Large Language Models (LLMs) exhibit behavior similar to human memory, but the underlying mechanism in LLMs has not been thoroughly explored.
- The paper uses the Universal Approximation Theorem (UAT) to explain LLMs' memory mechanism and conducts experiments to assess their memory abilities.
- Introduces the concept of "Schrödinger's memory," suggesting that an LLM's memory only becomes observable when queried.
- Comparisons between LLM memory and human memory highlight similarities and differences in operational mechanisms.
- Poems are dynamically generated by LLMs based on input, similar to how human memories are recalled through specific prompts.
- Both the brain and LLMs operate by dynamically fitting outputs based on inputs, indicating a shared fundamental mechanism of reasoning ability.
- Research extends this concept to other cognitive abilities such as social skills and creativity, attributing them to reasoning based on existing knowledge and inputs.
- Despite exhibiting reasoning capabilities and creativity aligned with linguistic conventions, LLMs may underperform in reasoning tasks due to factors like model size and data quality/quantity.
- LLMs have become integral tools impacting various fields like machine translation, text summarization, and sentiment analysis.

## Summary

Memory is a crucial aspect of human cognition and serves as the foundation for daily activities. With the rapid advancement of Large Language Models (LLMs), some models exhibit behavior similar to human memory. However, the underlying mechanism of memory in LLMs has not been thoroughly explored. This paper utilizes the Universal Approximation Theorem (UAT) to explain this mechanism and conducts experiments to assess LLMs' memory abilities. It introduces the concept of "Schrödinger's memory," suggesting that an LLM's memory only becomes observable when queried. By comparing LLM memory to human memory, similarities and differences in operational mechanisms are highlighted. The study delves into how poems are dynamically generated by LLMs based on input, similar to how human memories are recalled through specific prompts. It suggests that both the brain and LLMs operate by dynamically fitting outputs based on inputs, indicating a shared fundamental mechanism of reasoning ability. The research extends this concept to other cognitive abilities such as social skills and creativity, attributing them to the capacity for reasoning based on existing knowledge and inputs. Despite exhibiting reasoning capabilities and creativity in generating outputs aligned with linguistic conventions, LLMs may underperform in reasoning tasks due to factors like model size and data quality/quantity. However, they have become integral tools impacting various fields like machine translation, text summarization, and sentiment analysis. Understanding the intricate workings of memory in LLMs not only sheds light on their cognitive processes but also provides insights into the broader landscape of artificial intelligence research and its implications for society.