Which model to use
Every model stores and recalls the same memories. What differs is how much context comes back per query, what it costs, and which extras the engine can serve. Pick one per store and stay with it.
Compare models
| Feature | Tablet 2 | Scroll 1.2 | Book |
|---|---|---|---|
| Description | The current Tablet, at the Tablet price. Knows who said what, stores images, and can be asked again. | The current Scroll. Finds memory by the sentence, so a fact said once in passing still surfaces. | A different design, built so it does not repeat the same mistake twice. |
| Input | $2 / MTok | $4 / MTok | not priced yet |
| Output | $3 / MTok | $8 / MTok | not priced yet |
| Context per query | ~1,000 tokens | ~2,800 tokens | — |
| LongMemEval-S | 95.7% 93.7% read by a different frontier model | 92.3% 5-run mean, σ 0.4 | — |
| Images as memories | ✓ | — | — |
| Ask again (verify) | ✓ | — | — |
| Delivery forms | ✓ | ✓ | — |
| Engrams | ✓ | ✓ | — |
| Status | Live | Live | In development |
| Report | Tablet 2 | Scroll 1.2 | Book |
The table shows the current model in each line. Tablet 1 and Scroll 1 are still live and still supported — search above to put one back in the table.
Prices are per million tokens, plus a flat $0.0001 per request. Storage is free and there is no memory cap. Benchmark numbers are five-hundred-question runs; the readers and the judge are named in each model report. See Pricing and how we benchmark.
What the extras are
Images as memoriesStore a picture as the memory itself and find it from a sentence in any language. Tablet 2 and newer. Reference.
Ask again (verify)When one pass does not carry the answer, ask again and get memories it has not already returned. Tablet 2 and newer, on search only. Reference.
Delivery formsRender each recalled memory’s time in human language (memoir) or as a dated record (archive). Scroll 1.2 and up, and Tablet 2. Reference.
EngramsBuilt-in multi-hop retrieval pipelines with no language model in the loop. Tablet 1 and up. Reference.
A model whose engine does not implement an endpoint answers 501 rather than
charging for a call that did nothing.
Choosing one
- Start on Tablet 2. It is the current Tablet at the Tablet price, and it is the only model that stores images and can be asked again.
- Move to Scroll when one memory was never going to be the whole answer. Scroll returns a fuller context for questions where completeness matters more than cost.
- Pick one model per store. A store written through Tablet 2 and then read back with Tablet 1 or Scroll can return less than the same store written through those models.
Selecting a model on a call
Send the X-WOS-Model header, or set it once in the SDK client and override a
single call. Omit it and the account default is used. The live list is at List models, which needs no API key.
curl -X POST https://api.wontopos.com/api/v1/memory/recall \
-H "X-API-Key: $WOS_API_KEY" \
-H "X-WOS-Model: tablet-2" \
-H "Content-Type: application/json" \
-d '{"user_id":"alice","query":"what does alice drink?"}'Live models
- tablet-2 Tablet 2 The current Tablet, at the Tablet price. Knows who said what, stores images, and can be asked again.
- tablet-1 Tablet 1 Lean and fast. The lowest-cost way to give an agent long-term memory.
- scroll-1.2 Scroll 1.2 The current Scroll. Finds memory by the sentence, so a fact said once in passing still surfaces.
- scroll-1 Scroll 1 Reads wider and hands the reader a fuller context, for questions where completeness matters more than cost.