Perpetuity

Agent context isAgent context is

ExpensiveExpensive

The whole conversation history is re-sent every turn.The whole conversation history is re-sent every turn.Most of it is not needed for the current prompt.Most of it is not needed for the current prompt.

Every messageEvery message

KeptKept

For 99.1% of questions the evidence reaches the agent within one recall call, for 97.4% already in the first frame.For 99.1% of questions the evidence reaches the agent within one recall call, for 97.4% already in the first frame.Measured on the 470 questions that have evidence.Measured on the 470 questions that have evidence.

ReasoningReasoning

94.094.0

from a 3.6k-token framefrom a 3.6k-token frame

LongMemEval.LongMemEval.Opus 5 with thinking, the official GPT-4o judge, 500 questions, nothing excluded.Opus 5 with thinking, the official GPT-4o judge, 500 questions, nothing excluded.Half of them answered in one round.Half of them answered in one round.