Capstone project

Build an assistant that gets measurably better because it remembers, and prove it.

The brief

Create a personal or team assistant (for example a coding helper, a study buddy, or a customer-support agent) that keeps useful memory across sessions. It must handle changing facts correctly and let users inspect and delete what it knows.

Requirements

1. Short-term memory: persisted threads with a compaction policy, tested on a long task.
2. Long-term memory: at least two kinds (for example semantic profile plus episodic log) with a write path that consolidates instead of appending.
3. Retrieval: scored retrieval (similarity plus recency or importance) with a cap on injected memories.
4. Updates and history: superseded facts are marked, not silently lost.
5. User control: a page or command to list, edit and delete memories.
6. Evaluation: at least 25 test cases across recall, multi-session, temporal, update and abstention, compared against a full-context baseline.
7. Safety: owner isolation test and at least three poisoning attempts documented.
8. Write-up: architecture diagram, results table with accuracy, tokens and latency, and an honest "where it fails" section.

Suggested path

Stretch goals

• Add a reflection job that turns episodes into semantic facts.
• Model facts with validity intervals and answer "as of" questions.
• Store relationships in a graph and answer "who and what is connected" queries, building on the GraphRAG course.
• Compare your store against LangMem, Mem0 or Letta on the same test set.

Share it

Publish the repo and the results. Include failures: a clear account of where memory hurt (stale facts, wrong retrievals, extra cost) shows real engineering judgment.

Where to go next