Agent memory for LLMs — a self-pruning knowledge context graph modeled on human memory. Long-term, navigable, context-budget-aware memory with no vector DB, no embeddings, no server. Markdown + git. Claude Code plugin.

A feed-forward neural network that learns to prune itself during training using differentiable gate parameters

Self-pruning MLP-Mixer on CIFAR-10 every Linear is a custom PrunableLinear with a learnable sigmoid gate, trained with CE + λ·Σσ(g). 83.86 % accuracy, up to 128× compression, verified by hard-pruning every sub-threshold weight. Submitted for the Tredence AI Engineering Internship 2026 case study.