# Unified Latent-State Memory Fabric (UL-SMF)

A memory-compression system for long-context transformer inference that combines finite scalar quantization with dynamic latent-space mapping to shrink key-value cache tensors, with an open orchestration layer and a separately licensed core compression binary.

UL-SMF targets the memory bottleneck of long-context transformer inference by compressing key-value cache tensors using finite scalar quantization combined with dynamic 16-dimensional latent mapping through a component the project calls the oracle core. The README reports compression of up to 384 times while maintaining over 94% semantic retention in its own benchmarks.

The orchestration framework and interfaces are open source under AGPLv3, but the full production pipeline requires a compiled, separately licensed oracle-core binary for high-performance compression; the open-source quickstart script defines the pipeline's structure, while local execution of the full pipeline depends on that licensed component. Commercial licensees receive the optimized binary plus integration support and zero-copy VRAM routing.

Benchmarks in the README were run against an 8B-parameter reference model using a custom latent dynamic-cache class, reporting figures for semantic fidelity on a multi-needle retrieval test and measured VRAM footprint reduction. It is written in Python.

Website: https://github.com/liventruth/UL-SMF-Cache-Compression