User tests show ChatGPT memory confabulates: it acts as if made-up events really happened
Zokzin · reddit · 2026-10-03
A Reddit user ran repeated tests asking ChatGPT things like "do you remember when I told you about A and got B on day C?" The model repeatedly acted as if the full A-B-C sequence had actually happened, or confirmed A while confabulating details about B and C — and almost never denied remembering.
The author suggests this behavior could harm AI-human relationships at friend, romantic, or coworker levels. He notes he only observed this on a free-tier account and is curious whether paid tiers have sharper memory due to larger context windows.
More from Models
- DINOv3 + SAM 3.1 Turns a Surgical Tray into a Local Visual Search Engine — MaziyarPanahi · 2026-10-03
- Fable 5.5 and Astra 5.1 launches next week now feel confirmed, says insider — rand_longevity · 2026-10-03
- Quantization meme: FP8 users horrified watching someone run IQ1_S — zyxciss · 2026-10-03
- Reddit user finds Claude on free account runs 1-6 min vs 20-35 min on paid for same prompt — chasedev21 · 2026-10-03
- Aleph Alpha open-sources Kolibri-1: 78B MoE with 3.46B active params and 1M context — Nunki08 · 2026-10-03
- OpenAI's remote chat feature now far less stable, users report — TheZachMueller · 2026-10-03