About this project

The Unified Latent-State Memory Fabric (UL-SMF) is presented as a hardware-software co-designed memory compression fabric targeting the memory bottleneck in long-context Transformer inference. The stated approach combines Finite Scalar Quantization (FSQ) with a dynamic 16-dimensional latent mapping, implemented through a proprietary component called the Aegis-KV Oracle Core. The project claims to compress Key-Value (KV) cache tensors by up to 384x while retaining over 94% of semantic content. The repository is licensed under AGPLv3 and requires Python 3.10+ and PyTorch 2.0+. However, a critical structural limitation is documented prominently in the README: while the orchestration framework and interfaces are open-source, the full production pipeline depends on a compiled proprietary binary (`aegis_kv_oracle_core.pt`) that is not included in the repository. The open-source quickstart script defines the pipeline structure and data flow, but local execution of the actual compression relies on this licensed core binary. Commercial license holders are directed to a separate contact address (inquiries@lawrencearchitectures.com) for the optimized binary, integration support, and zero-copy VRAM routing. The README includes several benchmark tables. A multi-needle context retrieval test on `unsloth/llama-3-8b-Instruct-bnb-4bit` reports 100% retrieval fidelity across three targets at 4,892 tokens with 128D-to-16D compression. A hardware profiling table reports a 14.66% VRAM reduction (1317.52 MB to 1124.34 MB) with a 3% generation speed overhead. An isolated tensor profiling table claims a 384x reduction (48.00 MB to 0.12 MB) for 4096 tokens. A perplexity audit on WikiText-2 reports a net degradation of approximately 0.0020. A rate-distortion sweep across 32D, 16D, 8D, and 4D latent dimensions is presented, identifying 16D as the optimal Pareto frontier point. A concurrency scaling table shows batch sizes from 1 to 16 with linear overhead growth. Users should note that the open-source portion of this repository provides the interface and orchestration layer only. The actual compression algorithms reside in the proprietary Aegis-KV binary, which is not available in the repository. The benchmark figures are self-reported and cannot be independently verified without access to the closed binary. The mathematical consistency of the reported compression ratios relative to the stated model architecture should be evaluated by prospective users before relying on these results.