Skip to main content
sovseal gives the Claude Code CLI a persistent memory layer that works across terminal sessions, shell environments, and machine restarts. Because Claude Code supports stdio-based MCP servers natively, you can wire up the entire integration with a single command — no external signup required for the free local-only tier.

​
Installation

Run the following command in any terminal to register the sovseal MCP server with Claude Code:
This command downloads the @sovseal/mcp-server package and adds its stdio transport mapping to Claude Code’s configuration. The entry is saved to ~/.claude.json automatically.

​
Verify the installation

1

Start an interactive Claude Code session

Open a new terminal and run:
2

List available tools

Inside the interactive session, type:
You should see store_memory and recall_memory listed among the active tools.
3

Run the onboard validation prompt

In your Claude Code session, prompt Claude to store and recall a test fact:
Claude will execute store_memory followed by recall_memory, confirming the ONNX embedding pipeline warmed up and LanceDB local commits pass.

​
Configuration file location

Claude Code persists its MCP server registrations in a JSON file you can inspect or edit manually:
  • Location: ~/.claude.json
The sovseal entry looks like this:
~/.claude.json

​
How it works

​
Warmup and model cache

On first tool execution, the MCP server:
  1. Starts the embedding pipeline using the CPU-bound Xenova/all-MiniLM-L6-v2 model (~22 MB).
  2. Downloads the ONNX model parameters to ~/.sovseal/models/ if they are not already cached.
  3. Validates model integrity with SHA-256 pins (ensureAndVerifyModel) before accepting any requests.
After the initial download, all subsequent loads complete from disk cache in under 25 ms.

​
Memory storage

Every memory you store is written to a serverless LanceDB table at ~/.sovseal/db/memories.lance. No local database daemon is required. If you enable the write-behind replication worker, the client encrypts data with AES-256-GCM before any bytes leave your device, so the remote endpoint only ever stores ciphertext.
The cold start download (~22 MB) happens only once. After the model is cached, you will not notice any startup delay between sessions.