Open Claude’s memory settings and check what it has saved about your work. Correct outdated details, narrow preferences that apply only sometimes and remove information you no longer want reused.
This guide follows Anthropic’s documentation, checked on 8 September 2026. I haven’t carried out the audit in a live account for this piece. The examples are fictional.
Find the memory controls
In the current experience, open Settings > Memory and look under Topics. Open an entry to read it, then use the edit icon or Delete. If your account instead shows Memory under Settings > Capabilities, use its legacy View and edit memory control. Anthropic’s memory guide.
If you want a copy before editing
For a small audit, copy the entries you intend to change into a file you are permitted to keep. Treat that file as work information, choose its location deliberately and remove it when it is no longer needed. Compare any chat-generated memory summary with the entries you can see before treating it as a backup.
Temporary migration note, checked 8 September 2026: Anthropic says a legacy-memory export option in Settings > Memory is available until 9 September 2026. This is for recovering context lost in the migration. If it is visible and you need that copy, export before editing; if it has gone, continue the Topics review. The note does not set a deadline for ordinary memory housekeeping. Anthropic’s migration notice.
Check the entries most likely to affect an answer
Start with your employer, role and active projects. Then read writing preferences and repeated assumptions about how you work. You are looking for saved context that could make a future reply wrong.
For example, “I lead the Northbridge migration” might need to become “The Northbridge migration has ended; do not treat it as my current project.” That is an illustrative correction. Use the wording that is true for your own work.
You can also tell Claude in chat what to change or forget; Anthropic says those edits apply to the next conversation. After correcting an entry, reopen it and check the saved wording. A confirmation in chat is not enough reason to skip that check.
- Outdated: a completed project still appears to be your current assignment. Remove it or clearly mark it as past work.
- Too broad: “Always write informally” came from one social post. Narrow it to the situation where it applies.
- Incorrect: a colleague is listed as the decision-maker when that responsibility belongs elsewhere. Correct the role using information you know.
- Unnecessary: an entry includes personal or client detail that adds nothing to future answers. Remove it rather than keeping it by habit.
Keep current project facts in a maintained brief
My recommendation is to use memory for context you expect to reuse, such as your role and writing preferences. Keep dates, approvals and live status in a brief you can review against the source. A remembered deadline may have been accurate when you mentioned it.
For that separate task, use a standing brief in ChatGPT or Claude.
Check what deleting an entry changes
Removing a memory entry does not establish that the original information has disappeared from every record. Anthropic treats chat search and memory as separate controls, and says deleting a conversation does not remove related memory entries. Review both when you want to stop old context being reused.
Avoid a full reset for a few wrong entries. Anthropic says Reset memory permanently deletes memories, including project memories. Team and Enterprise owners should also note that switching memory off for the organisation deletes everyone’s existing memory. Memory controls and deletion behaviour.
Finish by opening a new conversation and asking Claude about one corrected detail. Check the answer, then return to Topics if it is still wrong. Repeat this audit after a role change or the end of a project, when old context is most likely to mislead.