The history doubles. The reading levels off.
Step 1 of 4: The history
A steady reading cost, however long the history.
LongMemEval asks questions about long chat histories between a person and an assistant. As the history grew to forty sessions, the reading behind each of Duplex’s answers levelled off, and where a fact had changed, it answered with the current one.
LongMemEval, by Wu et al., ICLR 2025. The benchmark →
- 96%Current answers where a fact had changedThe new fact, not the old one
- 3.7×Less reading per question than reading the whole historyAt forty sessions of history
- 84%Accuracy at forty sessions of history
3.7× less reading for every answer.
At forty sessions of history, against sending the whole history to the model.
Step 1 of 3: Reading
Duplex sent 24,176 tokens a question. Sending the whole history meant 88,798.
It answers with the fact as it stands now.
Where a fact changed partway through the history, Duplex gave the current answer 96% of the time.
- Temporal reasoning: 90% correct.
- Reasoning across several sessions: 89% correct.
Step 1 of 4: Stated
Early in the history, a fact is stated: the team uses vendor A.
How it was run.
- Benchmark
- LongMemEval, by Wu et al. (ICLR 2025), from the cleaned public release.
- The work
- Answer a question about a long chat history between a person and an assistant.
- Histories
- The sessions holding the answer, mixed with unrelated sessions from a pool of 19,195, at 2, 5, 10, 20 and 40 sessions.
- Scale
- 50 questions at five history sizes: 250 answers for every approach.
- Models
- GPT-5.5 read what each approach supplied and wrote every answer; a separate model judged them.
What it means for your work.
Long-running work stays affordable.
From twenty to forty sessions the history doubled, and the cost of each answer stayed at $0.032.
Answers follow what changed.
Where a fact had changed, Duplex answered with the current one 96% of the time.
It answers, rather than holding back.
Duplex declined 5 of the 250 questions, and answered the rest.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.