docs/analysis/turboquant-applicability-2026-03-31.md

TurboQuant Applicability

Source reviewed:

• src/skynet/doc/TurboQuant - Online Vector Quantization with Near-optimal Distortion Rate.txt

Best fit inside OpenSkyNet

• The strongest fit is not session authority, world model, or executive routing.
• The best fit is vector-state storage used by experimental memory systems such as src/omega/holographic-memory.ts.
• That area already stores normalized embeddings and computes cosine similarity over persisted vectors.

Why it fits

• TurboQuant is online and data-oblivious. That matches OpenSkyNet better than codebook-training approaches.
• holographic-memory is storage-heavy and experimental enough to tolerate approximate compression with bounded retrieval loss.
• It does not force benchmark/project logic into the authority chain.

What was implemented now

• Added a small, honest, TurboQuant-inspired path for normalized embedding compression in src/omega/embedding-quantization.ts.
• Integrated it into src/omega/holographic-memory.ts for newly stored fossils.
• Old fossils with dense embeddings remain readable.

What was deliberately not done

• No claim that this is the full TurboQuant algorithm from the paper.
• No attempt to quantize world-model, kernel, or session authority state.
• No new learned codebooks or calibration passes.

Recommended next steps

1. Benchmark cosine-recall impact on holographic-memory resonance over larger synthetic fossil sets.
2. If the retrieval loss stays low, evaluate the same approach for hierarchical-memory episodic z_state.
3. Keep all authority-chain state unquantized unless there is hard evidence that compression is needed there.