ForkKV: Scaling Multi-LoRA Agent Serving via Copy-on-Write Disaggregated KV Cache

ForkKV targets a specific bottleneck in the 'Agentic Era': when multiple LoRA-tuned agents process the same massive context (e.g., a codebase or legal document), traditional prefix caching fails because each LoRA's unique weights cause the hidden states (and thus KV caches) to diverge immediately. The project introduces a 'Copy-on-Write' disaggregated cache that maximizes sharing until divergence is mathematically necessary. While technically sophisticated, its defensibility is low (4) because it is essentially a high-performance optimization for the inference stack. Major serving frameworks like vLLM, S-LoRA, or Predibase's LoRAX are the natural gravity wells for this logic. With 0 stars and 3 forks, it currently exists as a research artifact rather than a production-grade tool. Frontier labs and infrastructure providers (Anyscale, Together AI) are highly likely to implement similar logic internally to reduce the TCO of multi-agent workflows. The displacement horizon is short (6 months) as this capability is a logical next step for existing block-manager-based inference engines.