Residual Vector-based Reconstruction as Long-Context Recall Regardless of Context Window Size

Large language models (LLMs) process long contexts, including long documents and lengthy conversations, but face token-level memory usage that increases proportionally to input length. Although model optimization and lossy prompt compression are widely used, these methods still fail to solve the lon