The surface we unrolled to write more. Scroll reads your question more carefully, then brings back more of what matters.
“It re-reads the question before it answers, so the right memory isn't missed.”
Indexes every sentence in its own right, so a thing said once in passing can be found by that sentence and still arrives with the passage around it. 92.3% on LongMemEval-S, about 2,800 tokens per query. Adds the Memoir and Archive delivery forms.
Full benchmark reportThe LLM-assisted tier: a frontier model joins the loop on our keys, so Scroll answers more completely than Tablet. Your stored memories are never sent to it, only the query. 90.7% on LongMemEval-S, about 3,700 tokens per query.
Full benchmark report