FLASHMEMORY DEEPSEEK-V4 RETRIEVER

Summary from the TLDR AI newsletter

FlashMemory predicts which DeepSeek-V4 CSA KV-cache chunks future tokens will attend to, keeping only the most relevant chunks on GPU. The retriever reportedly preserves or improves downstream performance while retaining about 10–15% of the KV cache on-device.