The abstract of the arXiv paper "Memorilla: Latent Semantic Memory for LLMs" starts from a familiar problem: large language models are limited to what is stored in their parameters. Memory, the authors note, is what allows an LLM to go beyond that, drawing on things like prior conversations, long documents, or an agent's earlier actions.

The paper's contribution is Memorilla, described as a latent semantic memory. The abstract specifically frames this in the context of a frozen LLM—one whose parameters are not being updated—so the memory mechanism has to supply accumulated knowledge externally rather than through retraining.

The source is only an abstract snippet, so the specific architecture, experiments, and results are not available here. The paper is listed on arXiv under identifier 2610.21375.