How well AI remembers · 97 questions, August 2024, GPT-4o
A reread beat the memory.
ChatGPT's memory, on the same short history.
A full reread of that history.
A small pilot, using the products as they were then. No independent test of today's ChatGPT memory was found. The bars share one scale: percent of the 97 questions. Wu et al., LongMemEval.
| Figure | Its own test | In the picture |
|---|---|---|
| 57.7% and 91.8% | LongMemEval, August 2024, 97 questions, GPT-4o. Memory, then a full reread. | Yes |
| 36.3 and 74.3 | COLLABMEM scores, not shares of questions. Letta with gpt-4o-mini: notes inside one role, then the same system plus a shared team archive. | No |
| 61% | A search over one shared store leaked protected content on 61% of access probes. The home page says these are different exams, and each score belongs to its own test. This leak is not a third result for the Letta pair. | No |
| 25.8% and 80.1% | GateMem medical summary, GPT-5.4. Mem0, then reading the history. | No |
| 57.0% and 56.5% | GateMem office setting. A search that carries the access policy, then reading the history. | No |
| 46.01% | GroupMemBench average accuracy for Hindsight. A separate exam of conversations among several people. | No |
| 47% | Gartner, 2022, published May 10, 2023. Digital workers who struggled to find the information they needed. 4,861 people in four countries. | No |
| 5.5% | US knowledge workers who meet Section's proficiency bar in its June 2026 report. The bar now includes agent and automation skills. 73.5% are experimenters. | No |
| 41% | CISOs, companies with more than 1,000 employees, who reported RAG over enterprise data in broad or limited production. a16z, 100 leaders, chart dated October 18, 2026. RAG is access. It is not evidence the organization can find, maintain, or forget on purpose. | No |
| 10% | McKinsey, fielded June 25 to July 29, 2025. In any one function, including knowledge management, no more than 10% said their organizations were scaling AI agents. | No |
Deleting the chat does not by itself remove a saved memory made from it, at Anthropic, Microsoft, and OpenAI. That sentence has no share in it. The section on the home page.