Eldric's forecaster predicts time series: it reads the recent past of a series, together with related series measured alongside it, and predicts the next steps as a band of 99 quantiles.
Eldric's memory and Eldric's forecaster share one simple idea. Each new item is written into a small grid as a key (where it goes) and a value (what it says). Later, a question reads the matching value back. Eldric remembers across sessions; the forecaster keeps its memory while it reads a series.
One slab per step. W/S or ↑/↓ walk · ←/→ step · drag sideways to look, up/down to walk · wheel zooms · hover a cell to see what changed.
Each slab is one step. The grid on it is the memory. Across: where a key points. Up: what a value says. Blue cells are positive, rose cells negative; taller and brighter means stronger. Cells lit in amber are the ones this step changed most.
Every item is stored as a pair: a key that says where it goes and a value that says what it is. To recall, you ask a question: the memory returns whatever was stored under keys that resemble it.
Each step writes a little and keeps most of what was there. Old items fade slowly unless they are written again, so the memory stays useful instead of filling up with stale detail.
Eldric's memory keeps what you give it across sessions, so Eldric remembers. The forecaster's memory lives while it reads one series: it collects what matters so far and uses it to forecast the next steps.
| Question | Eldric's memory | Eldric's forecaster |
|---|---|---|
| What is stored | Associations you give it: something to recognise, and what to recall for it. Kept across sessions. | What it has seen so far in the series and the related series alongside it, held while it reads them. |
| How it forgets | Slowly and evenly, at a pace you can set. Items fade unless they are refreshed. | At every step it decides how much of the past is still worth keeping. |
| How it is read | Ask with a question; get back the stored association that matches it best. | At each step it asks its memory what matters next, then forecasts the coming steps as a band of 99 quantiles. |
| What learns | Nothing needs training: it stores what you give it, immediately. | Trained on many series: how it writes, keeps and reads is learned from data. |
All numbers on this page are small illustrations computed in your browser. They are not Eldric's real sizes or trained values, and the grids are far smaller than the real ones.
Inspired by the interactive KronLSTM walkthrough (kron-lstm.vercel.app); no code reused.