loci
Does where a memory is filed decide what it can recall, and what it can imagine?
Pre-registered, two rounds · port of Vector-HaSH (Nature 2025)
0.009mean error
A theory with zero fitted parameters predicted recall on 156 held-out runs it was never developed on. Every run within 0.05.
The question
Vector-HaSH stores memories the way the hippocampus is thought to: a fixed grid-cell scaffold of 3,600 addresses. The paper leaves one choice open, where each memory goes. loci measures what that choice decides.
What I built
A theory of noisy recall with nothing fitted, frozen and hashed before it was tested; a filing rule derived from its error law (store where your recall points, then replay); and a test of whether the memory can construct combinations it never experienced.
What I found
The theory predicted 156 held-out runs to a mean error of 0.009. The filing rule beats k-means by 0.05–0.15. But filing for recall does not find the structure imagination needs: it constructs 0.24–0.31 of the unseen combinations, against 0.95–1.00 when filed by the true factors.
What it does not show
A study of how scaffold memories work and fail, not a better memory: kNN over the stored patterns recalls everything. The rule needs more state than the memory holds, and 2 of 13 pre-registered criteria failed.
- 0.009
- mean error over 156 held-out runs, zero fitted parameters
- 11 of 13
- pre-registered criteria passed as registered
- 0.24–0.31
- unseen combinations constructed when filed for recall
The result
| half the place cells | 0.004 |
|---|---|
| masked cues, 50% unknown | 0.004 |
| sentence embeddings | 0.007 |
| a fourth grid module | 0.007 |
| factor counts no module matches | 0.008 |
| other module periods | 0.009 |
| loads between the developed ones | 0.007 |
| cue dimension 2,000 | 0.010 |
| the paper's own write rule | 0.029 |
How it works
- noisy cue → cue map · ridge → place cells → grid modules 3 · 4 · 5 → address → read-outcoefficients · snap
- error law → filing rule → placementstore where recall points
- replay → placementexact descent
- read-out → filing through the read-out