Why the great theories of learning have never reached every learner — and what changes when the agent, not the assistant, becomes the unit of educational software.
For more than two thousand years, our deepest insights about how people learn have come from a small set of traditions: Socrates' dialogic method in 5th-century Athens, Piaget's constructivist account of cognitive development, Feynman's “teach-it-to-test-it” technique. They differ in vocabulary and era, but they converge on the same picture — the learner is an active constructor of knowledge; cognitive conflict drives change; understanding deepens through dialogue; every learner needs an interlocutor who knows them.
These ideas are universally admired. They are almost never implemented at scale. The reason is structural, not philosophical: each method requires a sustained one-to-one relationship with someone who knows the learner well, has unbounded patience, and can adapt in real time. Industrial education was built on the opposite assumption — one teacher, many students, fixed pace, standardized assessment. The result is a two-millennia gap between what we know about learning and what classrooms can deliver.
Agentic AI changes the constraint. A persistent agent — one that holds memory, intervenes on its own initiative, deepens its model of the learner over time, and runs at near-zero marginal cost — is the first technology in human history that can host the kind of relationship these theories require. It is not a faster textbook. It is the missing implementation layer. The three reports below trace the argument: from the shared structure of the classical theories, to the specific AI capabilities that unlock each one, to a concrete architecture for the agents that would carry them.
A three-part research arc — theory, practice, architecture. English versions below; the original Chinese is linked from each card.
Theory · Comparative Analysis
A structural analysis of three pedagogical traditions across two and a half millennia, identifying nine shared commitments — learner as active constructor, cognitive conflict as catalyst, metacognition, depth over coverage, individualization, dialogic interaction, simplification through analogy, the teacher as guide, and intrinsic motivation.
Practice · LLM Capabilities
Maps the seven structural barriers that have prevented these theories from scaling — teacher supply, time, assessment, cost, culture — onto specific capabilities of large language models. Argues AI is the first technology able to deliver high-quality, personalized pedagogical interaction without proportional human cost.
Architecture · Persistent Agents
A technical architecture paper. The Assistant paradigm — stateless, reactive, identity-less — cannot host the relationships these theories require. A Digital Daimon, built as a persistent agent with memory, autonomous intervention, deep learner understanding, and self-evolution, can. Six architectural properties, with their mapping to OpenClaw and to each pedagogical theory.
Education has always been a relationship problem disguised as a content problem. The work of the next decade is not to put more content in front of learners — it is to give every learner a persistent companion that knows them, asks the right question at the right time, and grows with them over years.