christopherkarani/Wax

Shared Single-file memory layer for all your agents, sub mili-second RAG over text, photo and video on Apple Silicon.. No Server. No API. One File. Pure Swift

What it solves

Wax provides a durable, local-first shared memory layer for AI agents and on-device models. It eliminates the need for hosted vector databases or manual markdown scratchpads by allowing multiple AI tools (like Claude Code, Cursor, and Hermes) to read and write to a single, searchable .wax file on disk.

How it works

Wax uses a single binary file to store documents, metadata, and indexes. It combines SQLite FTS5 for full-text search and CoreML-powered MiniLM embeddings for semantic vector search. To enable agents to use this memory, it provides an MCP (Model Context Protocol) server that acts as a broker, allowing different AI clients to share the same memory store via HTTP. For Apple ecosystem developers, it offers a Swift SDK and native integration for Apple Foundation Models, providing tools like waxRemember and waxRecall directly to the device's on-device LLM.

Who it’s for

  • AI Agent Users: People using tools like Cursor, Claude Code, or Grokbot who want their agents to remember facts across different sessions and projects.
  • Apple Developers: iOS and macOS developers building AI apps that need a durable, on-device memory store without relying on cloud APIs.
  • AI Researchers/Power Users: Those seeking a private, account-free alternative to hosted memory services.

Highlights

  • Local-First: All data is stored in one .wax file, which can be easily moved via iCloud or AirDrop.
  • Hybrid Search: Combines fast text search and semantic vector search (p95 recall ~6.1ms on Apple Silicon).
  • Cross-Agent Compatibility: A single memory store can be shared across multiple different AI agents via MCP.
  • On-Device Embeddings: Auto-configures MiniLM embedders on iOS 18 and macOS 15+.
  • Zero-Config Memory: Unlike some RAG systems, remember stores text directly without requiring an LLM to extract facts first.

Related

  • Project
    mcncarl/agent-memory-vault

    Agent Memory Vault is a Python toolkit that provides a local, Git‑backed Markdown vault with SQLite (and optional vector) search, enabling Claude Code, Codex and similar AI agents to share durable, auditable long‑term memory across sessions.

  • Project
    ThinkfleetAI/memmesh

    A persistent, local-first memory store for AI agents that provides durable, typed, and searchable memory using a single Rust binary without requiring a hosted vector database.

  • Project
    doobidoo/mcp-memory-service

    A self‑hosted memory backend for LLM agents that provides a REST API, CLI, OAuth, and a web dashboard. It stores facts, decisions, and causal knowledge graphs locally (SQLite + ONNX embeddings) or in hybrid/cloud mode, offering fast (≈5 ms) semantic search, tag‑scoped retrieval, and optional LLM‑generated summaries. Works with LangGraph, CrewAI, AutoGen, Claude Desktop, ChatGPT, and many other tools.

  • Project
    MemPalace/mempalace

    MemPalace is a local-first memory system for AI assistants (Claude Code, Codex, Cursor, any MCP client). It stores conversation histories verbatim — no summarization — and retrieves them via semantic search from a structured index organized into "wings" (people/projects), "rooms" (topics), and "drawers" (content). Pluggable vector backends (ChromaDB default; SQLite-exact, Rust-native, Milvus, Qdrant, pgvector optional) keep everything on-device by default, with no API calls needed. It ships a CLI, a 47-tool MCP server, a temporal knowledge graph, auto-save hooks, and Docker images. Claims 96.6% raw retrieval recall (R@5) on LongMemEval without any LLM, with reproducible benchmarks and unusually honest methodology notes.

  • Project
    Intuition-Lab/personal-model

    A local-first long-term memory runtime for macOS that creates an evidence-linked personal model of a user's work patterns and goals to provide grounded context for AI coding agents via MCP.