If a careful reader went through your whole AI history — every conversation, months of it — they’d come away with a set of real patterns about how you work. We built that baseline: 12 patterns from 38 real conversations. Then we blind-matched what Lethe extracts against it. Today’s product surfaces 5 of the 12. That’s 41.7%, and we publish it next to our 98.1% precision because one number without the other is marketing.
Where the misses live
Every single miss is a cross-conversation pattern. Nobody ever says “I work in intense bursts and then go quiet for days” in any one chat. The sentence doesn’t exist. The pattern lives in the gaps between sessions — in when things were said, not in anything that was said.
Extraction reads one conversation at a time, so a pattern that only exists between conversations is invisible to it by construction. No better prompt fixes that. It’s not a model problem; it’s an architecture fact.
What closes the gap
Two things, both measured. First, timing is arithmetic, not inference: gaps, cadence and clustering computed in code from timestamps we already store — that alone moved recall from 41.7% to 58.3%, and it cannot hallucinate, because no model is involved. Second, a synthesis pass over already-stored claims, run in chunks (one pass over everything returned nothing — a silent prompt-size failure we also published). The full union reaches 11 of 12: 91.7%.
Honesty about status: the 91.7% is a test harness, not the shipped product. What ships today is 41.7% plus the deterministic rhythm facts. When the union pass ships, this essay gets a successor with the shipped number — that’s the deal this site makes with you.