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

The task

An education platform needed long-term memory of each student: not a gradebook or prompt context, but a model of topic mastery, forgetting and current cognitive state. Existing approaches don't give a persistent neural model of a student, and the memory dynamics had to provably converge. We framed it as research to be taken all the way to a service the platform can integrate.

What we built

• We built a ~23.7k-line Python research platform. Its neuromorphic memory core — 59 modules in 12 subsystems — runs on spiking neural networks; the architecture is derived automatically rather than hand-designed — details in patent prep, closed.
• We created a cognitive digital twin of a student: it tracks mastery and forgetting, estimates cognitive state, predicts performance and picks the right task difficulty.
• We formally verified convergence of the memory core in Lean 4 with Mathlib: 22 theorems and lemmas, zero unproven statements, machine-checked.
• We added orchestration of heterogeneous models — 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. Three consumers in the platform's main backend are wired to it.

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

22

Lean 4 theorems and lemmas on memory convergence, zero unproven statements, compiler-checked

32

REST endpoints, 12 of them the digital-twin API; counted from router code

More cases