RAG without the stack: one .wax file on Apple Silicon

What it does

Wax is a Swift-native memory engine that stores an AI agent’s documents, embeddings, and structured knowledge inside one self-contained .wax file. It exposes a Swift API, a CLI, and an MCP server so coding assistants like Claude Code can recall context across sessions. Everything stays on device; there are no cloud dependencies, Docker containers, or API keys to manage.

The interesting bit

Instead of splitting text and vector search across separate services, Wax bundles SQLite FTS5 and a Metal-accelerated HNSW index into the same binary container. That makes the whole database as portable as a PDF—you can AirDrop it, back it up, or sync it via iCloud.

Key highlights

• Hybrid search (text + vector) with a p95 recall latency of roughly 6 ms and cold-open times around 9 ms on Apple Silicon, per the project’s benchmarks.
• Single-file architecture: documents, metadata, and both search engines live in one frame-based .wax container.
• Pure Swift, targeting iOS and macOS; runs offline and leverages Metal for vector indexing.
• Ships with an MCP server so agents can remember facts and recall context across sessions.
• Ingest throughput is claimed at ~86 docs/s with full hybrid indexing on an M3 Max.

Caveats

• Vector search requires an embedder; if it is unavailable, hybrid or vector queries fail loudly rather than falling back to text-only mode.
• Apple-ecosystem only: iOS and macOS on Apple Silicon, so don’t expect it to run on your Linux server.
• Benchmarks were run on an M3 Max; performance on other Apple Silicon chips is unspecified.

Verdict

Wax is worth a look if you are building on-device AI features for macOS or iOS and want to kill the network round-trip. If you need cross-platform vector search or already run a managed cloud RAG pipeline, it is not for you.