Long-term student memory: from hypothesis to a working service

The task

An education platform needed long-term memory of each student: a model of topic mastery, forgetting and current state. A gradebook or prompt context does not give that. We framed it as research: build such a memory and take it to a service the platform can plug in.

What we built

• We built a research platform of about 23.7k lines of Python. Its neuromorphic memory core, 59 modules in 12 subsystems, runs on spiking neural networks; the architecture is derived from a compact genome rather than designed by hand. The code is open.
• We built a student model: it tracks mastery and forgetting, estimates current state, predicts performance and picks task difficulty.
• We added formal proofs in Lean 4 with Mathlib: 21 statements about the convergence of the classical Hopfield network, with no gaps (zero sorry). This re-formalises a known result; the spiking core is not covered by these proofs.
• We added orchestration of different model types (SNN, DNN, CNN, RNN, transformers) with 7 execution strategies and hybrid architectures.
• We packaged it all as a FastAPI microservice: 32 endpoints, health checks, a Docker image, Redis and RabbitMQ integration. The platform’s main backend has three connection points for it, but the client is switched off in production: this research line is not used in the products.

Outcome

10 days

from first commit to a release with formal verification, measured by commit dates

368

test functions across 37 files, counted from code

21

Lean 4 statements about the classical Hopfield network, zero sorry, compiler-checked; the spiking core is not covered

32

REST endpoints, 12 of them for the student model; counted from router code

Services

More cases