Every number, losses included
Each figure below is generated from a recorded run and lives in
paper/neurips/generated/numbers.tex. Nothing here is hand-typed.
Continual stream, 100 sequential tasks
Immediate accuracy is measured at the moment of learning; retained accuracy after the full stream. Chance is 0.0625.
| System | Immediate | Retained | What it isolates |
|---|---|---|---|
| Page memory (Hebb) | 1.000 | 1.000 | Allocation and a write |
| Episodic store | 0.988 | 0.988 | Allocation, no write |
| Fast weights | 0.667 | 0.072 | A write, no allocation |
Full fine-tuning and LoRA sit with fast weights, for the same reason: one shared state, overwritten.
Allocation predicts retention
Varying only the page count across a 128× range (1–128 pages) leaves immediate accuracy at 1.000 throughout, while retained accuracy moves from 0.068 to 1.000. The fraction of memories still holding a page of their own predicts retained accuracy to within 0.0039 across all 6 settings, with no free parameters.
Addressing — the open problem
Over 198 rules, learned content addressing sits on a frontier rather than clearing it.
| Measure | Value |
|---|---|
| Best false-merge rate | 0.353 |
| Best missed-revisit rate | 0.222 |
| Best retrieval@1 | 0.717 |
| Key separation, meta-trained | 0.897 |
| Key separation, untrained | 0.425 |
Where we lose
On vision, a plain episodic store beats page memory outright: 0.994 against our 0.697. That is on the home page footer for the same reason it is here.
| Benchmark | Page memory | Episodic |
|---|---|---|
| Vision | 0.697 | 0.994 |
| Compositional | 0.458 | 0.374 |
Footprint
Per memory, against an episodic store on the same stream — not against a per-tenant adapter, which is a different comparison.
| Store | Bytes per memory |
|---|---|
| Page memory | 1,164 |
| Episodic | 56,448 |
| Ratio | 48× |