How it's evaluated
OpenAI evaluates memory on three axes across three generations (2024 saved memories → 2025 + Dreaming V0 → 2026 Dreaming V3) and reports the lift as charts. The article publishes no citable percentages, so this page invents none — it shows the directional claim. The one concrete figure: serving compute cut ~5× for Free users.
Serving compute, Free (approx.)≈5 × lower
Axes of “good memory”3
Generations compared3 2024–2026
Published % figures0 citable
| Evaluation axis | 2024 · Saved | 2025 · + V0 | 2026 · V3 |
|---|---|---|---|
| Carrying forward context (factual recall) | Baseline | Improved | Best |
| Following preferences (preference adherence) | Baseline | Improved | Best |
| Staying current (correct over time) | Baseline | Improved | Substantial lift |