The professor's handoff
Three studies Prof. Wu ran before handing the work over, with numbers copied from RESEARCH_HANDOFF_CONTINUOUS_MEMORY.md (branch DPFWM). We did not rerun them, except the Amazon part of the first, which our rerun reproduces.
The professor's grid versus network test
This is the study our rerun copies. Each model tried three learning rates and three temperatures for ten epochs; seed 42 picked the best on tuning users. The plan then froze those choices and scored three seeds on held-out users, but the handoff was written with 53 of 54 tuning runs done and nothing held-out yet. Numbers here are copied from the handoff, not rerun. Its final results are not in Git, and his lab folder isn't readable from our account.
Both network versions beat the tuned grid on Amazon tuning users; normalized and additive readouts tied.
Amazon tuning NDCG@10 · best at each temperature
| Model | T 0.1 | T 0.2 | T 0.4 |
|---|---|---|---|
| Grid memory (FWPRec) | 0.264987 | 0.258699 | 0.247028 |
| Network, normalized | 0.277522 | 0.268444 | 0.256640 |
| Network, additive | 0.277037 | 0.277452 | 0.272666 |
Best settings
| Model | Learning rate | Temperature | Epoch |
|---|---|---|---|
| Grid memory (FWPRec) | 0.003 | 0.1 | 9 |
| Network, normalized | 0.003 | 0.1 | 9 |
| Network, additive | 0.003 | 0.2 | 10 |
EBNeRD
Only the grid memory had finished: 0.516832 tuning NDCG@10 (lr 0.003, T 0.2, epoch 8). The network grids were incomplete, so there is nothing to compare.
Source: RESEARCH_HANDOFF_CONTINUOUS_MEMORY.md, branch DPFWM, 19 Oct 2026. Tuning data only. Status: Partial.
The professor's earlier three-seed check
An earlier matched study with the same data split and the same shared starting profile, but a smaller tuning budget: learning rate only, no temperature search. It is where the shared starting profiles in his later study came from. Held-out user scores were averaged over three seeds. The setup differs from the full-grid study, so these numbers should not be mixed with it.
A small Amazon gain for the network and a tie on EBNeRD, with a lighter tuning budget.
Held-out users · mean NDCG@10 over 3 seeds
| Dataset | Grid memory (FWPRec) | Network |
|---|---|---|
| Amazon | 0.25693 | 0.26287 |
| EBNeRD | ≈ 0.51314 | ≈ 0.51317 |
Source: RESEARCH_HANDOFF_CONTINUOUS_MEMORY.md, which gives the EBNeRD values as approximate. Results folder: results/associative_memory_confirmation/selection42_seeds42_43_44_v1/ on his machine. Status: Complete.
Can the memory explain itself?
Concept memory routes each write through named concepts so a score can be broken into contributions and edited locally. Mechanically that worked. As a recommender it was far behind the continuous network. In this pilot the additive readout also looked ahead of normalized, which is what the later grid study tested with temperature tuning, and found to be a tie.
Explainable concept memory cost more than half the ranking quality. The direction moved to continuous memory.
Held-out users · NDCG@10 · seed 42
Source: RESEARCH_HANDOFF_CONTINUOUS_MEMORY.md; results/concept_memory_pilot/amazon_seed42_v1/ on his machine. One seed, single pilot. A related VQ codebook memory was also tried and stayed less accurate after stabilizing; the handoff gives no scores for it. Status: Complete.