Can GPT-4 perform tasks requiring evolving long-term memory and context?

Ascertain whether the large language model GPT-4 can perform tasks that require an evolving long-term memory and context, such as reading a full-length book and following the plot across chapters, including understanding references to earlier material.

Background

In the concluding section, the authors outline several limitations that need to be addressed for progress toward more general AI. One key limitation they highlight is the absence of long-term memory in GPT-4, as the model operates with a fixed context and does not maintain state across extended interactions.

Within this discussion, they state that it is not even clear whether the model can handle tasks requiring evolving memory and context, using reading a book and tracking plot references across chapters as a concrete example. This uncertainty motivates investigation into capabilities and potential architectural or tooling changes needed to support such tasks.