Building offline-first AI features in React Native (New Architecture) usually means stitching together MMKV for persistence, Zustand for reactive state, and custom native bridges for vector search or token streaming.
I just shipped v3.1.0 of react-native-ultrastore to unify all three into a single, zero-dependency, synchronous JSI runtime built on MMKV v4 + Nitro Modules.
What’s Inside (36.1 kB, 0 Dependencies)
- Synchronous JSI Storage & State: Reactive hooks (
useUltraStore), atomicsetManybatching, encrypted namespaces, and strict schema migrations with zero async bridge overhead. - On-Device Vector Search (
react-native-ultrastore/ai): Cosine similarity search operating directly on cachedFloat32Arraybuffers without re-validation overhead—ideal for local RAG/embeddings under ~50k vectors. - Bounded LLM Stream Cache (
UltraStreamCache): High-frequency token buffering with automatic memory cleanup and snapshot caching so streaming AI responses don't choke React renders. - Production Hardening: Automatic recovery for corrupted disk payloads, stale query cancellation in
useUltraQuery, and 42 lifecycle tests. - AI-Agent Ready: Ships with root
llms.txt,llms-full.txt, and.cursorrulesso Cursor and Claude generate accurate, idiomatic code out of the box.
Quick Example
tsx import { createUltraStore, useUltraStore } from 'react-native-ultrastore'; import { VectorStore, UltraStreamCache } from 'react-native-ultrastore/ai'; // 1. Synchronous MMKV v4 store const store = createUltraStore({ id: 'app-store' }); // 2. On-device vector similarity search const vectors = new VectorStore({ dimensions: 384 }); await vectors.add('doc-1', new Float32Array(384), { title: 'Offline RAG' }); const matches = await vectors.search(queryEmbedding, { topK: 5 }); // 3. Bounded LLM token stream cache const streamCache = new UltraStreamCache({ maxTokens: 4096 }); streamCache.append('msg-1', 'Hello '); 
