Memory-augmented large language models must decide which memories to retain, and recent systems do so by estimating each memory's effect on task performance. However, as the abstract of a new arXiv paper notes, these estimates rely entirely on retrieved memories. The paper argues that this reliance creates a problem, though the abstract cuts off before specifying the full consequences.
The proposed solution, called a Causal Memory Policy, intervenes on retrieval to make memory utility identifiable. By explicitly intervening rather than passively observing retrieved memories, the method aims to provide a clearer causal link between a memory and its contribution to task performance.
The source is a single arXiv abstract, so the details of the intervention and experimental results are not yet available from this text. The paper's significance lies in identifying a potential blind spot in current memory-augmented LLM design and offering a causal framing to address it.