Shipped: LSE Mnemosyne

Mnemosyne is a self-hosted AI team memory and corporate wiki. Notes live as markdown files in your own git repository. Any MCP-capable AI (Claude, Cursor, and others) can search, read and write them through nine MCP tools, with hybrid full-text and semantic search in one SQLite file. A web portal then turns the notes of each project into documentation: an AI-written article, a PDF handbook and architecture diagrams painted by an image model. Nothing calls an AI model until an admin agrees to the estimated cost.

Who it is for: engineering teams and companies that want their AI assistants to share one memory, and their documentation to write itself, without sending notes to a SaaS. People sign in with GitLab and see only the projects they belong to.

The number: in one run it documented 92 projects for $0.51 in text-model cost. The image model is counted separately, per picture.

Open source (Apache-2.0), built during the Meta Global AI Developer Hackathon with Meta Muse models. Free for up to 4 people. Every install gets a 30-day trial with everything unlocked.

After that, 5 or more seats need a licence at $99 per seat per year, with a 10% discount from 5 to 50 seats.