AI Tutor Always Forgets? Someone Built an MCP Server That Stores Learning Progress in SQLite

After a conversation with an AI tutor, it may still have to ask what you know from scratch next time. The learning state isn't stored separately, so every conversation feels like starting over.

In a tweet, Dan Kornas introduced a project called Tutor MCP. It is an MCP server for developers and self-learners, and its goal is to bring structured, personalized tutoring into AI assistants. MCP can be understood as a protocol that lets AI assistants connect to external tools and data; in this case, it connects a tutoring engine to the AI client you commonly use.

The core capabilities include several parts:

• Course graph: breaks a subject into concepts, prerequisite relationships, and goals, avoiding just giving a loose list of topics.
• Knowledge tracing: uses Bayesian Knowledge Tracing to estimate what you have mastered from recorded interactions.
• Spaced repetition: uses FSRS to schedule reviews so that content learned earlier stays in the learning loop.
• Session memory: saves summaries, goals, and recurring misunderstandings for use in the next conversation.
• Local learning history: runs via stdio and uses SQLite for storage; compatible clients can share history through the same installation.

The project uses the MIT license and is currently marked as alpha. The GitHub repository is here, and the original tweet is here.

What's worth looking at in this project is that it leaves explanation and practice to the AI while handing progress management to an independent engine. The chat interface is still there, but learning history no longer lives only in the chat window. For self-learners, the difficult part is often knowing what to practice next and when to go back and review. Tutor MCP tries to make these two things engine capabilities.

Someone in the replies asked whether learning history is separated by subject or all content shares one file. The original post did not elaborate on this. Based on the available information, it can only be confirmed that it uses SQLite to store local learning history, and compatible clients can share history through the same installation.