ContextVault: Stop repeating yourself to your LLMs
I've been tracking this new tool called ContextVault. It’s essentially a shared memory layer designed to solve exactly this. Instead of having your project knowledge scattered across individual LLM "Projects" or local files, it acts as a centralized vault that any AI client can tap into.
The clever part is the MCP (Model Context Protocol) server integration. If you're using an AI client that supports MCP, you don't have to manually copy-paste instructions. You can literally tell the model to "save this to the vault" or ask "have we handled this specific edge case before?" and it searches the database via the server to pull the relevant context. It’s built on a serious stack: PostgreSQL with pgvector for the semantic search, Node.js, and TypeScript. It even handles OAuth for GitHub and GitLab, so permissions aren't just a suggestion—they're baked into the workflow.
I looked into the architecture, and they're using Clerk for auth and strictly scoped organization storage. This isn't just a toy for solo devs; it’s built for teams that need role-based access control so a junior dev isn't accidentally pulling sensitive architectural secrets they shouldn't see.
If you're tired of the "copy-paste-repeat" cycle, you might want to check out how they've structured their MCP implementation. Here is the core concept of how you'd interact with a system like this via a prompt to ensure the model uses the vault:
Retrieve the architectural decision record regarding our PostgreSQL indexing strategy from the ContextVault and apply those constraints to the following code snippet.It's a solid move toward making AI an actual team member rather than just a series of isolated, amnesiac chat windows.
https://www.contextvault.dev