SAE Feature-Triggered Subspace Weight Consolidation for Large Language Models with Full Memory Lifecycle Dynamics

We present a memory consolidation system for large language models (LLMs) implementing a full knowledge lifecycle: continuous fact extraction from dialogue, context-sensitive SAE-triggered retrieval, offline subspace weight consolidation, and structured forgetting via time-based strength decay. The system uses GemmaScope-2 sparse autoencoder (SAE) features at layer 16 of Gemma-3-4B as memory retrieval triggers, and performs constrained least-squares updates to MLP down-projection weights at layers 25–27 with Gram-Schmidt orthogonalization to protect consolidated knowledge. Evaluated on 1,000 CounterFact cases, the method achieves 78.0% efficacy, 77.6% generality, and 100% specificity, with a 29× smaller weight perturbation than standard fine-tuning (ΔW = 0.004 vs. 0.118). Mechanistic analysis confirms that rank reduction is localized to layers 25–27, with a 3.4× edit-specificity ratio. The system also reproduces Ebbinghaus-style forgetting dynamics and demonstrates parametric implicit memory persistence after index clearance. This is a preliminary preprint (v1.0, April 2026). Code and updated versions will be released in future revisions.