← All systems
OpenAI Memory (ChatGPT)
OrganizationOpenAI (platform)
Founded2024–2026
Funding—
Codeclosed
Saved memories + full-history reference; 2025 teardown shows a generated dossier injected at chat start, not RAG.
Every number on file
| claimed | benchmark | config | runner | status |
|---|---|---|---|---|
| 41.5 → 82.8 | "its own evals" | no dataset, no harness (Dreaming, 2026-06) | self (via press) | UNTRACED |
| 52.90 | LoCoMo (J) | measured by a competitor | Mem0 paper | THIRD-PARTY |
0 of 6 major platforms published a reproducible memory benchmark at launch — all rigorous numbers come from third parties.
Rows are not comparable — configs differ. A row graduates to a verdict only through a registered re-run. The protocol →