I fitted the new δ-mem research for apple silicon using mlx and openclaw integration! My findings

So I’ve been nerding out hard about memory, and have come to the conclusion that context management is too high level and dynamically changing the weights would be best. Luckily, this morning I checked my news feed and saw this new paper! https://arxiv.org/abs/2605.12357 It improves model attention