The idea
FWPRec keeps a short-term memory of each user as two grids of numbers. The professor's handoff asks whether a tiny neural network, trained one step at a time on that user's clicks and skips, makes a better private memory. Everything else stays fixed, so any change in ranking comes from the memory.
What changes, what stays
The memory
Two 64 × 64 grids become one two-layer network (hidden width 32). Each event triggers one gradient step on that network instead of a grid correction.
Everything else
Frozen Qwen item vectors, the long-term profile u₀, the last-20-event window, the candidate lists, the loss and the metrics.
The question
Does the network memory rank better than FWPRec when both get exactly the same tuning?
Does scoring with the raw query (additive) beat the length-normalized query once the score scale is tuned?
Self-explaining memories (concept and codebook versions). Accuracy was poor; paused until the main question is settled.
What not to claim
Neural memory, test-time adaptation and concept explanations all have prior work. A sourced novelty check is still owed.
The memory is rebuilt from the last 20 events for every request. No network is kept for a user's whole history.
All results so far use development data that earlier pilots had seen. No test split has been scored.