scalingfactbearishLinear-attention and state-space language models suffer from lossy memory where earlier facts are overwritten when many key-value associations competeArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02303v1