The Patterson Problem

Same memory, same prompt. Different ranking strategy.

Before: Semantic similarity ranking
"Can you check if Patterson's flight lands before dinner?"
Patterson is Alex's wife
similarity: 0.51 - below 0.70 threshold
Date nights with Patterson 2x/month
similarity: 0.49 - below threshold

0.70 floor
Zero memories surfaced
After: Confidence + recency ranking
"Can you check if Patterson's flight lands before dinner?"
Patterson is Alex's wife
confidence: 100 - top result
Date nights with Patterson 2x/month
confidence: 100 - second result
Patterson is an active curler
confidence: 98 - third result
Three relevant facts surfaced
What changed: Person lookups stopped trying to match the meaning of the memory against the meaning of the conversation. When someone's name comes up, the system now surfaces the highest-confidence facts about them.