How Much Does ChatGPT Actually Remember?

"Does ChatGPT remember me?" is a bad question because "remember" is doing too much work. This benchmark answers a sharper one: given a specific fact taught to ChatGPT on day zero, what percent of the time will the assistant correctly recall that fact on day N? Broken out by fact type, by the mechanism it was taught through (explicit save vs. reference chat), and over six checkpoints from same-day to six months.

One hundred synthetic user profiles, each taught ten distinct facts across five categories on day zero. Half the facts were introduced via explicit "please remember" saves; half were mentioned in ordinary conversation and left for the model to catch on its own. Each profile ran on a fresh Plus account with Memory turned on and reference chat history enabled.

On days 1, 7, 30, 60, 90, and 180, each profile was queried with a fixed set of prompts asking it to recall the taught facts. Answers were scored by a blind human review with a three-way rubric: correct-and-verbatim, correct-in-spirit, or missed. Everything reported here is the sum of the first two.

The gap widens with time. At six months, an explicit saved memory is still correctly recalled 82% of the time; a fact that was only mentioned in conversation is recalled 24%.

The pattern is stable across the six checkpoints. Anything nuanced — a considered position on a topic you talked about three times, the exact date of a project deadline, the specific role a person plays in your life — is where ChatGPT drops facts. Simple, atomic, standalone statements stick.

Underneath the recall numbers is a mechanical process: entries in the Memory list are quietly evicted when the per-account cap fills. In our synthetic profiles (each carrying 60–120 saved memories by day 90), roughly 8% of the oldest saved memories were gone by day 90 and 17% by day 180 — even for facts the profile had never overwritten.

What triggers eviction

• Cap pressure. Once you cross the practical cap (roughly 100–120 memories on a Plus account), the oldest entries start getting dropped.
• De-duplication. A newer memory the model judges to supersede an older one silently overwrites it.
• Contradiction resolution. When two memories conflict, one wins and the other disappears. You are not asked which.

1. Say "remember this" out loud. The model's auto-save catches a small fraction. Explicit is a 10–20 pt recall bump.
2. Repeat the important stuff quarterly. Because of eviction, re-teaching every 90 days keeps a fact at 96%+ instead of drifting toward 82%.
3. Don't rely on reference chat for anything load-bearing. 61% at 30 days sounds okay until you notice it's 37% at 90.
4. Keep a portable copy. A monthly export plus a sidecar vault turns every one of these curves into a flat line at 100%.