# RecallM: An Adaptable Memory Mechanism with Temporal Understanding for Large Language Models

**Authors:** Brandon Kynoch, Hugo Latapie, Dwane van der Sluis  
**arXiv:** [2307.02738](https://arxiv.org/abs/2307.02738)  
**Date:** 2023-07-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.02738v3/

## Key points

- Large Language Models (LLMs) have shown remarkable capabilities in Artificial Intelligence.
- Enhancing LLMs with long-term memory is crucial for developing Artificial General Intelligence (AGI) systems.
- RecallM is a novel architecture designed to provide LLMs with adaptable and updatable long-term memory.
- RecallM excels in belief updating and maintaining temporal understanding of knowledge.
- Extensive experiments demonstrate the effectiveness of RecallM, which is four times more effective than using a vector database for updating stored knowledge.
- The paper "RecallM: An Adaptable Memory Mechanism with Temporal Understanding for Large Language Models" by Brandon Kynoch, Hugo Latapie, and Dwane van der Sluis provides further details on the proposed architecture.
- The authors have made their code publicly available on GitHub.
- RecallM presents a promising approach to augmenting LLMs with long-term memory capabilities, enabling enhanced reasoning abilities and continuous learning through user interactions.

## Summary

, , , , 

In the field of Artificial Intelligence, Large Language Models (LLMs) have made significant advancements and demonstrated remarkable capabilities across various tasks and domains. However, as we strive to develop Artificial General Intelligence (AGI) systems, it becomes crucial to enhance LLMs with long-term memory to overcome limitations in context window and enable sustained reasoning, cumulative learning, and long-term user interaction. To address this need, we propose RecallM, a novel architecture designed to provide LLMs with an adaptable and updatable long-term memory mechanism. Unlike previous methods, RecallM excels in belief updating and maintaining a temporal understanding of the knowledge it receives. Through extensive experiments, we demonstrate the effectiveness of this architecture. Furthermore, our experiments on temporal understanding and belief updating reveal that RecallM is four times more effective than using a vector database for updating previously stored knowledge in long-term memory. The paper titled "RecallM: An Adaptable Memory Mechanism with Temporal Understanding for Large Language Models" by Brandon Kynoch, Hugo Latapie, and Dwane van der Sluis provides further details on the proposed architecture. The authors also make their code publicly available on GitHub for reference. Overall, RecallM presents a promising approach to augmenting LLMs with long-term memory capabilities. This paves the way for advancements in AGI systems by enabling enhanced reasoning abilities and facilitating continuous learning through interactions with users.