KH-1: recall facts read 1M+ tokens earlier in constant GPU memory. Archival-memory adapters for frozen LFM2 and Qwen3.5 models: code, eval logs, paper.
memory pytorch recurrent-neural-networks adapters language-models gemma state-space-models delta-rule liquid-ai linear-attention llm long-context qwen lfm2 babilong
-
Updated
Oct 18, 2026 - Python