A paper, Memorizing Documents with Guidance in Large Language Models,
written by Bumjin Park and Jaesik Choi is accepted at IJCAI-2024.

1. Problem

This work tackles the problem of storing document-wise memories in LLMs.

2. Proposed Method

This work proposes a document representation as a key to select memories.

In addition, we propose document guidance loss to encourage the disentanglement of document memories.

3. Theory

We link the metric spaces of document, representation, and memory selections.