When to deliver more

The question is not which tier is better

Both tiers share one retrieval core and differ only in delivery. Tablet 1 hands the reader a single-pass, bounded context. Scroll 1 expands the same retrieval hits with their session neighbors and delivers roughly 1.5× as many tokens, for about 2.2× the cost per query. So the useful question is not which is more accurate in the abstract. It is whether the extra delivery buys accuracy at a given memory size.

Small memory: it does not pay

On LoCoMo, where each question carries about 20K tokens of history, we ran 1,986 questions under one shared reader. Tablet 1 answered 76.5% correctly and Scroll 1 76.7%: three questions apart out of 1,986, with a confidence interval comfortably spanning zero. The tiers are statistically indistinguishable, and Scroll 1 costs 2.19× more. At this record length retrieval already surfaces what the reader needs, so the extra session neighbors are redundant weight. Tablet 1 is the efficient choice.

Large memory: it does, and we say how firmly

At about 140K tokens, on the 500-question benchmark we published earlier, Scroll 1 leads by 5.5 points (90.7% vs 85.2%). The gain is not spread evenly: it lands on preference inference (+11.3), multi-session synthesis (+10.1), and temporal reasoning (+6.3), exactly the questions where the answer is assembled from pieces laid down at different times. The two single-session categories stay within half a point, because a fuller delivery has nothing left to add when the answer already sits in one place.

But that study used a different reader for each tier, so the 5.5 points cannot be cleanly attributed to delivery alone. An earlier delivery ablation points the same way, and the category structure is consistent with it, yet the honest statement is this: the small-memory side is a settled, same-reader result, and the large-memory side is directionally supported but reader-confounded.

Memory beats reading the whole record, on both axes

At about 102K tokens per question, we also measured what happens if you skip memory and hand the reader the entire record. It is not just more expensive. It is also less accurate.

The cost gap widens with every query

A record is usually asked about more than once. Memory ingests the record once and then retrieves a bounded context; reading the full history pays to re-read everything on every single query. Memory is cheaper from the very first question, and the gap compounds from there.

The rule

The practical rule is not "Scroll is better." It is this: deliver more only when the record is large enough that fuller delivery buys accuracy, and stop paying for it when it does not. Where the gain is zero, pay less and choose Tablet 1. Where the record is long and the work is integration-heavy, Scroll 1's extra delivery earns its extra cost. That is why both tiers exist.

What we have not proven

We state the limits with the same weight as the results. The new measurements are single-run, so run-to-run variation is not quantified here. The one well-powered, same-reader comparison we ran is a tie. The 24-question pilot at 102K is too small to separate the tiers. The large-memory advantage carries a reader confound. And two benchmarks at two sizes are two operating points, not a curve: proving the crossover needs a same-reader rerun at 140K, a controlled length sweep on one benchmark, and multi-run repetition.

One more thing to say plainly: these are our own products, and this report was written by the person who builds them. That is a conflict of interest. We offset it by publishing every per-question score, every prompt, and the full cost ledger, reporting no best-of numbers, and keeping a third-party re-verification path open. We treat the crossover as a hypothesis we find compelling but have not yet proven, and we invite the replication that would confirm or refute it.