Persistent memory for AI coding tools. An MCP server that lets Claude Code, Cursor, and Windsurf retain decisions, patterns, and project context across sessions. Local-first, open source, two-minute setup.
Every Claude Code session starts from zero. The agent forgets what you built last time, what you decided, and what didn't work. You reload context manually every single session. After running a fleet of agents as my workforce for six months, this was the most expensive friction in my workflow.
Context Vault is a local-first MCP server. It sits between any AI coding tool (Claude Code, Cursor, Windsurf) and a SQLite database with semantic search via sqlite-vec. The agent saves decisions, patterns, and project context once; on every subsequent session, it retrieves them automatically when relevant.
The design is opinionated about staying out of the way: install with one command, no cloud account, no telemetry. Your data lives in a local file you can grep, back up, or delete. The package is published to npm at v3.0 with 693 passing tests, and a hosted SaaS layer (rate limiting, billing, web dashboard) is also live for teams that want shared vaults.
Context Vault homepage at context-vault.com.
The product covers the full stack: an MCP server that integrates with Claude Code, Cursor, and Windsurf out of the box; a CLI for direct interaction; a published npm package; a SaaS dashboard for paid users; and ongoing canon for what to save and how (insight, pattern, decision, reference). I wrote about the architecture in A context vault for AI agents on the blog.
