Ryanne's blog

Experiments in systems and machine learning, written up as interconnected articles. Posts share a common vocabulary of topic pages, and every load-bearing number in a post is recomputed from code in this repository when the site builds — claims carry their evidence with them.

Posts

• Finite Memory for Unbounded Streams — what happens when you replace a transformer's exact attention with a finite memory: a two-tier KV cache with Nerode merging that generates coherently at 8× its trained context, beats the standard eviction baseline at a smaller budget, and — fine-tuned with the memory in the loop — approaches native-context quality at 8× compression. With a re-implemented inference engine, training pipeline, and a verified map of the KV-cache compression literature.