The pain: every session starts from zero
You explained the whole project to Claude last week. Today, it remembers none of it. So you paste the same context again. Then again in ChatGPT. Then again in Cursor. Every model upgrade resets the relationship.
The problem isn't the AI. It's that your knowledge isn't anywhere the AI can actually read.
What an MCP memory layer is
MCP (Model Context Protocol) is an open standard that lets AI clients pull live context from external sources through a single, well-typed interface. Instead of a vendor's tiny built-in "memory" feature, MCP gives any compliant AI a way to read a real knowledge base you own.
A memory layer over MCP means: one knowledge base, many AIs, no copy-paste.
How BrainTube does it
- Auto-capture. BrainTube ingests from YouTube (segmented transcripts), podcasts and audio, PDFs and EPUBs, web articles, Notion (4-hour sync), meeting audio, and screenshots.
- Compile. Each item gets semantic embeddings and entity extraction. A Postgres knowledge graph auto-links concepts; PageRank-style retrieval surfaces what matters. FSRS-5 spaced repetition keeps the important bits sharp.
- Query from any AI. A live MCP server (listed in the official MCP registry and on Smithery) exposes your library to Claude, ChatGPT, Cursor and any MCP-capable client.
The differentiator: memory that fills itself
Almost every "AI memory" or "second brain" tool is a note app you have to type into. BrainTube is the opposite shape: the memory grows from your existing consumption — the videos you already watch, the podcasts you already listen to, the PDFs you already open.
Tools change. Models cycle. Your mind's index stays.
How to connect
The fastest path is the one-click OAuth connector inside Claude.ai. For Cursor, Claude Desktop, and other clients, paste the streamable-HTTP config with your X-BrainTube-Token header. Full instructions are on the connect page, and the per-client setup guide walks you through each one.