Covalence - Cross-Client Memory for Claude, Cursor, and MCP

TL;DR

TL;DR: Covalence is a macOS menu bar app that acts as a shared MCP memory server for Claude, Cursor, and any MCP-capable AI tool — so context stored in one client is instantly available in all others, entirely offline.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: covalence.app
• Source repository: github.com/darronz/covalence (main app), github.com/darronz/covalence-website (marketing site)
• License: Unknown (repo not publicly accessible at time of writing)
• HN launch thread: news.ycombinator.com/item?id=47794883
• Release: covalence.app/releases/Covalence-latest.dmg

What Is Covalence?

If you use multiple AI coding tools daily — Claude Desktop, Claude Code, Cursor, or any MCP client — you have likely hit this wall: context you carefully explained in one session vanishes the moment you open another tool. Covalence solves this by running a local MCP server that acts as a shared memory bank for every AI client on your Mac.

“I use Claude Desktop, Claude Code, and Cursor daily. They all have memory now, but none of them share it. Something I explained in Claude Desktop doesn’t exist when I open Cursor.”

The author, darronz, built Covalence out of frustration with existing solutions that required Docker, external databases, or cloud accounts — all overkill for a personal context store.

Covalence is a native macOS menu bar app. You install it, configure your AI clients to connect via MCP, and done. Every client shares the same SQLite-backed memory store.

How It Works

Architecture

Covalence bundles three components into a single download:

• MCP server — any MCP-capable AI client connects to it as a tool server
• SQLite + sqlite-vec — local vector database stores semantic embeddings
• nomic-embed-text-v1.5 via CoreML — embeddings computed entirely on-device, no API keys, no network calls

Memory Storage

Each MCP client runs as a separate connection. The author solved the concurrency problem by using SQLite in WAL (Write-Ahead Logging) mode, which allows multiple readers and writers safely without deadlocks.

Retrieval

Covalence uses hybrid BM25 + vector search with recency weighting. When you ask for something, it finds what you mean, not just what you typed. You can pin your most important memories so they always surface first.

Setup Workflow

Step 1: Download and Install

Download the latest release from the Covalence website:

open https://covalence.app/releases/Covalence-latest.dmg

Mount the DMG and drag Covalence to your Applications folder.

Step 2: Add the MCP Server to Your AI Client

Open Covalence. The app displays an MCP snippet tailored to your setup. For Claude Desktop, add it to your claude_desktop_config.json:

{
  "mcpServers": {
    "covalence": {
      "command": "npx",
      "args": ["-y", "@covalence/mcp-server"]
    }
  }
}

The exact snippet varies by client — Covalence generates the correct one for your setup. Paste it into your AI client’s MCP configuration file.

Step 3: Paste the Behavioral Prompt

Covalence also provides a behavioral prompt that tells your AI when to store and search memories. Paste this into your AI client’s system prompt or instructions.

Step 4: Start Using It

That’s it. Store a memory in Claude Code, retrieve it in Cursor. Both connect to the same Covalence instance.

Key Features

• Semantic search — finds what you mean, not just keyword matches. Hybrid vector + BM25 with recency weighting
• Core Memories — pin the most important knowledge; always surfaced first across every client
• Cross-client sync — Claude Desktop, Claude Code, Cursor, and any MCP-capable agent share the same memory simultaneously
• On-device embeddings — nomic-embed-text-v1.5 runs via CoreML on your hardware. No API keys, no per-call cost, no data leaves your machine
• Menu bar app — one-click capture, global hotkey opens search from any window
• Data ownership — everything lives in a single SQLite file on hardware you own. Export as markdown or JSON anytime. No accounts, no telemetry

Under the Hood

Retrieval pipeline:
1. Query → embed with nomic-embed-text-v1.5 (CoreML, local)
2. Hybrid search: sqlite-vec (vector) + BM25 (keyword)
3. Results ranked with recency weighting
4. Return top-K memories to the AI client

The hard part was concurrency. Multiple AI sessions writing to the same MCP server would deadlock on a normal SQLite setup. WAL mode lets each MCP client hold a separate connection safely.

FAQ

Q: Does Covalence require an internet connection? A: No. All embeddings run on-device via CoreML. There are no API calls, no external dependencies, and no data transmitted anywhere.

Q: What AI clients does Covalence support? A: Any MCP-capable client — Claude Desktop, Claude Code, Cursor, and any other tool that implements the MCP specification.

Q: Is there a cloud or hosted version? A: No. Covalence is a local-only macOS app. Your data never leaves your machine.

Q: How is it different from just using Claude’s built-in memory? A: Built-in memory is siloed per-client. Covalence is shared across all your AI tools simultaneously. What you teach Claude Code is immediately available in Cursor — without re-explaining.

Q: What does the setup require beyond the DMG? A: Just the MCP configuration snippet that Covalence generates for you. No Docker, no database setup, no cloud account.

Conclusion

Covalence fills a real gap in the AI toolchain: persistent, shared context across multiple clients. If you bounce between Claude Desktop, Claude Code, and Cursor — or any combination of MCP agents — you know the pain of re-explaining your stack every time you switch tools.

The on-device CoreML embedding approach is the right call — no API keys, no per-call cost, no privacy concerns. The SQLite WAL concurrency fix is also elegant: the hard problem of multi-client writes is solved without a separate server process.

Download at covalence.app/releases.