v4.1: Add gentle memory decay (1%/day) and pruning (30 days stale)

Fix: Lower default similarity_threshold to 0.10 for hash embeddings

v4: SLM-inspired architecture with three-tier memory, neural links, self-tuning

v3 TUNED: All bottleneck fixes + optimized threshold (0.45)

Update mnemo.py with smart injection and real embeddings

Upload mnemo.py with huggingface_hub