Personalized large language models often rely on a retriever to pull candidate memories from a user's history. But not all of those candidates can fit into the model's limited context. A new arXiv preprint, titled "Recall Begets Recall," focuses on the memory ranking stage that decides which candidates actually get in, so the final answer is grounded in the most relevant parts of the user's past.
The abstract describes memory ranking as the stage that determines which retrieved memories enter the context. The title suggests a recurrent mechanism: the memories already chosen for the context help guide the selection of further memories, rather than ranking all candidates in a single pass. This iterative approach could allow the model to build a more coherent and useful set of memories.
Because the source is only the abstract, the article does not include specific performance numbers or comparisons to existing ranking methods. The contribution appears to be the recurrent framing itself, which treats memory ranking as a sequential decision problem rather than a one-shot scoring task. The full paper would presumably explain the architecture and evaluation in more detail.