An H33 Company

The cache engine behind
2,216,488 auth/sec.

Cachee is the post-quantum cache infrastructure that powers H33's production pipeline. CacheeLFU admission control, L0 hot tier at 30 nanoseconds, 512 KiB constant memory footprint. The reason the Substrate can sign 2.2 million attestations per second is that the ZKP lookup path runs through Cachee.

30 ns
L0 Read Latency
512 KiB
Constant Memory
1.87%
Pipeline Overhead
389,000x
vs Raw ZKP
What Cachee does

Two-tier cache with admission control

Cachee is not Redis with a different name. It is an in-process, two-tier cache engine with a CacheeLFU admission sketch that decides which keys deserve promotion to the L0 hot tier. The L0 tier is a 64-shard lock-free read path that serves lookups in 30 nanoseconds. The L1 tier is a DashMap concurrent hash map that handles misses. The admission sketch is a Count-Min Sketch at 512 KiB constant memory — the same 512 KiB whether you're tracking 100,000 keys or 10 million.

No TCP. No serialization overhead. No connection pool contention. At 96 concurrent workers doing tight FHE loops, TCP-based cache access caused an 11x throughput regression. Cachee runs in-process — the lookup is a pointer dereference through a shard, not a network round-trip.

Where Cachee runs in the H33 pipeline
// H33 production pipeline (per 32-user batch) FHE batch verify 937 µs // 83% of pipeline Cachee ZKP lookup 1.9 µs // 0.2% — 32 lookups × 0.060 µs each Dilithium attest 189 µs // 17% — SHA3 + sign + verify ──────── Total 1,128 µs // 35 µs per auth // Without Cachee: raw ZKP = 23,534 µs per lookup // With Cachee L0: 0.060 µs per lookup // Speedup: 389,000x
Cachee vs Redis at 96 workers
MetricCachee (in-process)Redis / TCP RESP
Lookup latency0.060 µs~50 µs
Pipeline throughput2,216,488 auth/sec136,670 auth/sec
Pipeline overhead1.87%11x regression
Connection contentionNone (in-process)96 connections competing
Admission sketch memory512 KiB (constant)N/A
Eviction policyCacheeLFULRU / LFU / volatile
Cachee products

CacheeEngine

The core two-tier cache with CacheeLFU admission, L0 hot tier, and DashMap L1. In-process. Zero network overhead. What powers H33's 2.2M auth/sec pipeline.

CacheeLFU

Admission sketch that decides which keys get promoted. Count-Min Sketch at 512 KiB constant memory. 1,239x less memory than DashMap at 10M keys.

Solana PnL Solver

22,857x speedup on Solana wallet PnL reconstruction. 6,814ms cold to 0.298ms warm. Open source at github.com/H33ai-postquantum/solana-pnl-cachee.

Cachee RESP Server

Drop-in Redis replacement with CacheeLFU admission. RESP protocol compatible. For workloads that need network access — not recommended for hot paths at 96+ workers.

Install
# Rust (recommended — in-process, zero network overhead) cargo add cachee-core --features cms-admission # Homebrew (CLI tools) brew tap H33ai-postquantum/tap && brew install h33 # Docker (RESP server — Redis drop-in) docker pull h33ai/cachee-server:latest docker run -p 6380:6380 h33ai/cachee-server:latest

Built by H33. Powering H33.

Cachee is an H33 company. The same engineering team that builds the Substrate builds the cache that makes it fast.