Show HN: A 6M-token movable window on a single 46GB GPU
2 days ago
- A novel approach freezes a language model and grows a persistent memory of verified solutions alongside it, avoiding retraining.
- Once a problem family is solved and independently verified, every new instance is answered with zero generation tokens, bit-exact, and deterministically.
- Across 180 instances spanning nine problem families and four architectures, the system achieves 180/180 accuracy at zero tokens per answer.
- Memory selection takes 1.4 microseconds, and full reuse completes in 6-23 ms at 36 mWh, with exact addressing eliminating errors (0 errors vs 94.3% for approximate retrieval).
- The store functions as a working context of up to 6 million tokens on a single 46 GB GPU, surpassing limits of engines like vLLM and SGLang.
- Frontier models outperform on raw reasoning, but for verified problem families, verified reuse costs zero tokens and returns identical bits every time, inverting the comparison.