How to Correct or Remove an Outdated Memory in ChatGPT and Claude

October 2, 2026 · 3 min read

First, sort out what you are removing, because three layers are involved. Saved memory is what the assistant has distilled and stored about you — a preference, a project fact, a name. The original source is the conversation, file, or connected account where it first appeared. Later output is what the model says in future replies, and it can echo a stale fact even after the memory entry is gone, if the source still exists or the model leans on old context. Fixing one layer does not fix the others.

For ChatGPT, OpenAI's help documentation puts the controls under Settings, then Personalization, then Memory — exactly what you see varies by account, plan, region, and platform. The summary there is a partial view; memory does not preserve every detail, so an absent entry is not proof nothing was stored. Removing a saved memory and removing its source are separate actions: relevant chats, Library files, and connected sources may each need their own handling, and deletion can take time. A model reply confirming deletion is not a backend-deletion audit.

For Claude, Anthropic's documentation describes a new experience built around Topics under Settings, then Memory, with edit and delete controls on each. Deleting a conversation does not remove related saved memory entries — those get handled in the Topics list. Pausing memory retains everything while stopping its use and creation; resetting removes memory permanently. Legacy accounts that predate Topics behave differently and are documented separately, so if your screen shows the older summary-based layout, check that section before acting.

These panels do not tell you whether the fix worked, so treat the following as a proposed check of observed behavior, not a test of erasure. Write down the specific stale fact. In a fresh conversation with memory enabled, ask a neutral question that would surface it without naming it — not "do you remember X?", but "what should I keep in mind when we work on this project?" If the stale fact appears, that is an observed repeated output, not proof a particular memory entry is still active; leftover source context or plain generation could produce it too. A clean answer only means the fact did not surface in that trial — it does not rule out future influence or confirm removal. Then handle the source layer with whatever your account exposes, typically by deleting the relevant conversation or file, and run the probe once more in a second fresh session.

As the publisher, one note: SentX's public terms do not establish the same per-memory controls. Its privacy policy says submissions are not confidential and may enter training or shared memory, and distinguishes account-level deletion from data already in training, model weights, or shared memory — a statement of published terms, not a verification of internal erasure.

A simple way to keep stale entries from piling up is a periodic pass: open the memory panel on the assistants you actually use, skim for entries that no longer fit, edit or delete them, and spot-check one in a fresh conversation.