Curate What ChatGPT Remembers — Don't Let It Accumulate

A habit: treat ChatGPT's Memory as a small, deliberate profile you maintain — not a junk drawer it fills automatically. You periodically review, edit, and delete stored memories.

Memory is injected into every chat. Stale or wrong facts (an old job title, a project you finished, a preference you changed) silently bias every future answer. A tight, correct memory makes ChatGPT feel like it knows you; a bloated one makes it confidently out of date.

Review memory whenever your context changes — new role, new stack, new project — and once a month regardless. Add memories on purpose for durable facts (your tech stack, tone preferences, constraints).

Don't store anything sensitive or short-lived. One-off task context belongs in the prompt, not in permanent memory, and secrets shouldn't be there at all.

Ask directly:
"What do you currently remember about me? List each memory item."

Then curate in Settings → Personalization → Memory: delete anything stale, and add durable ones explicitly:
"Remember that I write in British English and prefer answers without preamble."

A developer who switched from Python to Go found ChatGPT kept giving Python examples. Asking it to list memories surfaced a year-old 'works primarily in Python' entry — deleting it fixed the bias in one step.

Keep memory intentionally minimal and push richer, per-domain context into Projects instead. Memory = who you are; Projects = what you're working on. Separating them stops one project's context leaking into unrelated chats.

• Letting ChatGPT auto-save everything, then wondering why answers assume outdated facts.
• Storing secrets or client-confidential details in a feature designed to resurface them everywhere.
• Putting temporary task context in memory instead of the prompt, polluting future unrelated chats.