ProveKV Squashes Multi Agent LLM Memory By Up To 68x Without Losing Quality

ProveKV is a new open-source release that slashes the memory footprint of multi-agent language model systems by up to 68 times. It stores the shared part of a conversation once in a compressed pool, then gives each agent only its unique tail, delivering zero perplexity regression on a real 1.7 billion parameter model. The project provides detailed benchmarks showing 36x lossless and 68x lossy compression compared to an uncompressed f32 KV cache.

RecursiveIntell built ProveKV as a two-tier, content-addressed cache pool that combines FibQuant and TurboQuant codecs. They validated the system on SmolLM2-1.7B with WikiText-2, achieving bit-exact output quality while cutting storage dramatically. The repository includes reproducible scripts, audit gates, and receipts to verify every claim.

Two-tier pool architecture

Developers experimenting with multi-agent LLM setups on consumer GPUs can drastically cut the storage needed for shared context. Privacy-conscious professionals benefit from running collaborative agents locally without exhausting video memory. Small teams can prototype complex multi-agent workflows that previously required expensive hardware.

Known limitations and open work

The system does not reduce the live GPU cache size during inference, as it must decompress back to f32 before patching the model’s memory. Current decode wall-clock tests reveal the batched path is slower than per-vector processing, so no speed gain is promised. The team plans to extend validation to longer contexts, larger models, and diverse datasets beyond WikiText-2.

“Result: 36× lossless / 68× lossy memory reduction vs. f32‑raw KV cache on SmolLM2‑1.7B + WikiText‑2 (0% ΔPPL).” — Source: Reddit