Deadline-Tracking Study Agent with Long-Term Memory for University Students

The problem: deadlines live in six different places

A typical student's deadlines are scattered across a learning management system, course group chats, emails from teachers, a timetable image and notices from the department office. Scholarship and internship deadlines add more. Students miss submissions not because they are lazy but because nothing pulls these sources together, understands them and remembers what matters to them personally.

General chat assistants can answer questions, but they do not watch your sources, they forget or misremember details, and they cannot be trusted to act on your calendar without checking with you.

What already exists and the gap

• Google Classroom, Moodle and Canvas send notifications, but only for their own course posts, and many teachers post deadlines elsewhere.
• Google Calendar, Notion and Todoist hold deadlines well once you type them in yourself.
• ChatGPT, Gemini and Claude now offer memory features and scheduled tasks, but they are general purpose, do not read your course portal, and give you limited control over what is remembered.

The gap is an agent built for students that gathers deadlines from their actual sources, remembers the right facts per student, lets them correct its memory, and never acts without approval.

Inspiration

The MemGPT paper from UC Berkeley researchers (2023) treated the context window like a computer's RAM and external storage like disk, with the model moving information between them. It became the open source Letta framework. It is a good starting point for thinking about what belongs in context, what goes into long-term memory and how the agent retrieves it.

What you build

• Students: connect sources, see extracted deadlines with a link to the original notice, ask questions such as "what is due before Friday?", review a weekly plan and approve calendar changes.
• Memory screen: everything the agent remembers about the student, editable and deletable.
• Agent: extraction, planning and reminders, with approval steps for anything that changes a calendar or sends a message.

The technically interesting part

• Structured extraction. Turn messy notices into typed deadline records with Pydantic or Zod, nullable fields when the date is unclear, and a retry loop with validation errors.
• Memory design. Separate facts (courses, preferences), events (past deadlines and outcomes) and procedures (how this student likes to plan). Decide what is saved, when it expires and how corrections override old memories.
• Scoping and privacy. Every memory is keyed by student and every read is filtered in code. Nothing sensitive is stored, and data export and deletion work.
• State and checkpoints. Weekly planning pauses for approval and resumes, even after a restart.
• Evals. Test extraction on real-style notices, test that memory recall is correct after weeks of updates, and count stale or wrong memories.

How to evaluate it

• Extraction precision and recall on a labelled set of 50 or more notices.
• Memory accuracy: questions whose answers depend on facts from earlier weeks, including facts the student corrected.
• Deadlines missed by pilot students compared with their previous semester, as they report it.
• Cost per student per week and latency of common questions.

Suggested plan

• Weeks 1 to 2: interview 10 students, collect sample notices with permission, design the deadline schema.
• Weeks 3 to 5: connectors and extraction with validation and evals.
• Weeks 6 to 8: long-term memory, the memory screen and per-student scoping.
• Weeks 9 to 11: weekly planning, approvals, calendar integration and checkpoints.
• Weeks 12 to 15: three-week pilot, fixes, final evals and report.

Path to a product

Universities and student societies could offer this to every student, and coaching centres to their learners. A freemium student app is another route, as long as costs per student stay low, which your cost measurements will show. The strongest version integrates with one popular learning management system very well before adding others.

Scope warning

Use test accounts and notices shared with permission, never scraped from other people's accounts. Support one learning management system and one calendar in the core scope. Do not let the agent submit assignments or email teachers on its own.

Learn what you need on Sefism

• AI agents and MCP roadmap: the memory and state stage, plus structured output, frameworks and evals.
• AI agents cheat sheet: short-term and long-term memory, checkpoints, compaction and human approval.
• LangGraph Documentation: persistence, interrupts and long-term memory stores.
• Pydantic Documentation or Zod Documentation: for the deadline schema and validation.
• pgvector: storing and searching memories inside PostgreSQL.
• Langfuse Documentation: tracing each extraction and planning run.

Get help with this idea

The risk in this project is a clever demo that remembers the wrong things. Tauseef can help you design the memory model, plan the pilot and build evals that show your agent is genuinely reliable. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session