Intentional memory, shared across every AI client

Your memory, on machines you own. Never anyone else's cloud.

Version 1.4.0 · macOS 15+ · ~265 MB

Two copy-pastes. That's it.

Add the MCP snippet. Paste the behavioural prompt. Done.

Connect

Add the MCP server to your AI client's config. Covalence generates the snippet for you.

Claude Desktop — config.json

{
  "mcpServers": {
    "covalence": {
      "command": "/Applications/Covalence.app/Contents/MacOS/cov-mcp"
    }
  }
}

Instruct

Paste the behavioral prompt so your AI knows when to store and search memories.

Claude Desktop — Project instructions

You have access to Covalence, a persistent
memory layer. Search memory at the start
of conversations. Store important decisions,
context, and preferences. Use retrieved
context naturally …

Everything your AI needs to remember

🔍Semantic search

Finds what you mean, not just what you typed. Hybrid vector + keyword search with recency weighting.

⭐Core Memories

Pin your most important knowledge. Always surfaced first across every client.

🔗One memory, every client

Claude Desktop, Claude Code, Cursor, and any MCP-capable agent. Same memory, simultaneously, across all of them.

⚡On-device embeddings

Embeddings computed on your hardware, not someone else's. No API keys, no per-call cost, no external dependencies.

📌Always running

Menu bar app with one-click capture. Quick capture from anywhere. Global hotkey opens search from any window.

🔒Your data, your infrastructure

A SQLite file on hardware you own. Export anytime as markdown or JSON. No accounts, no telemetry, no one else's cloud.

Under the hood

How retrieval works

Every memory becomes a 256-dimension vector on hardware you control. Vector similarity meets keyword match via RRF. No third-party cloud, by architecture.

Covalence embeds each memory with nomic-embed-text-v1.5 via MLTensor on Apple Silicon, truncates the output to 256 dimensions, and stores it in SQLite alongside an FTS5 full-text index. Searches hit both indexes in parallel — vector similarity and BM25 keyword scoring — and merge the results with Reciprocal Rank Fusion at k=60. A hyperbolic recency factor then nudges newer memories upward, weighted at 10% of the final score so recency never overrides genuine relevance.

Full technical deep-dive →

Your data, your infrastructure

Covalence runs on machines you control. No accounts with us. No telemetry. No third-party cloud, by architecture.

Every component — the database, the embedding model, the search engine — runs on your hardware. No network calls to us, no API keys to buy, no external dependencies to trust. Your memories are a SQLite file you own, back up, and export whenever you want.