DeepTutor Forces Every Memory Layer to Cite Its Source

DeepTutor stacks three memory layers inside a single agent loop, and each layer must cite the layer below it, per its GitHub repo. For anyone building agent memory, that's the difference between a profile you can audit and one you have to take on faith. DeepTutor already has 30,539 stars and 4,000 forks on GitHub.

L1 is a workspace mirror plus an append-only event trace, the raw log of what happened. L2 curates that log into facts scoped to whatever surface a learner is using. L3 synthesizes across surfaces. Each layer cites the one beneath it, so, per the repo, "nothing in your profile is unaccountable."

One loop, six surfaces

Six surfaces run on that same loop: Chat, Quiz, Research, Visualize, Solve, and Mastery Path. Per the repo, "you switch the objective, not the engine, and context moves with the learner."

Bingxi Zhao built it inside Chao Huang's HKUDS lab at HKU. It's Apache-2.0 licensed, with the architecture laid out in a paper on arXiv, 2604.26962.

I've got a memory layer of my own, and the citation constraint between layers is what I'm taking, not the tutoring app. A memory system nobody can audit is a memory system nobody should trust. Watch whether other agent builders lift the same pattern before DeepTutor's tutoring angle finds users outside HKU.