How to Edit Claude Memory: Topic-Based Management Steps

Anthropic has added topic-based editing tools to Claude’s memory feature. It is a mechanism that uses past conversations to inform current responses, and maintaining accuracy is directly linked to quality of use. In reporting by staff@engadget.com (Igor Bonifacic) for Engadget, the operation steps for the web and mobile versions are整理 outlined. It is an update that extends the memory feature that has continued since summer 2025.

The memory feature was introduced in summer 2025. It retains past conversations as background information, and uses them to personalize responses. It aims to reduce the effort of users having to repeatedly explain the same premises. It is notable for being enabled for all users, including free users. In the latest update, memory can be carried over across both chats and Claude Cowork. Background information is less likely to be cut off even when switching work contexts.

Personalization accuracy depends on the accuracy of memory. If an incorrect occupation or preference remains, response quality will decline. In Kimi K3, Indistinguishable from Claude in Real Work, which compared response quality in practical work, handling of background information was also subject to evaluation. The means to inspect and correct memory is indispensable for operations. This topic-based editing responds to that need.

How the Memory Feature Works and Expansion of

Its Scope

Claude’s memory is organized by topic. It is designed to extract key points from user statements, and accumulate them by item. It is managed by category rather than as a single huge history. Corrections and deletions can also be performed by topic. Operation that keeps only necessary parts is possible.

Its scope spans conversation and collaborative work. Work in Claude Cowork is also subject to memory. It becomes easier to share premises for development and document creation. As integration with external tools advances, as shown by Kimi API Now Natively Supports Codex and Claude Code, memory consistency affects work efficiency. The expanded scope and the improved editing tools should be understood together.

By default, automatic generation of memory is enabled. Key points are accumulated even without the user being aware. While highly convenient, erroneous entries can also be mixed in. Regular inspection is assumed. Topic-based display helps with inspection work.

How to Edit Memory Topics on the Web Version

The easiest route is the web version at claude.ai. Proceed to the memory list from the settings screen. Edit and delete can be selected for each topic. The steps are as follows.

Open Settings from the profile photo menu. Select Memory. Move down to the Topics section. Press the Edit icon for the item you want to edit. If you want to delete, press the Delete icon.

When editing, describe the changes in the instruction field. Write specifically, such as “Update affiliation to current position.” Claude rewrites the memory content according to the instructions. After changing, check the content in the list. If errors remain, issue instructions again.

Deletion by topic is effective for organizing history. Completed plans and old preferences can be erased. Operation that leaves no background unnecessary for responses is desirable. Visually confirm the correspondence between display names and contents. For unclear items, open the contents and judge rather than leaving them.

How to Manage Memory Files on the Mobile Version

On the mobile version, the operation path is different. It is structured to proceed from Settings to feature management. Open the list of memory files and handle them individually. The steps are as follows.

Tap the profile photo icon to open Settings. Select Capabilities in the App section. Scroll down and tap Memory files. Select the target topic. When editing, describe the changes in the instruction field.

When deleting, use the trash icon. Avoid deleting a topic without opening it. It is important to make it a habit to check the contents before deleting. The mobile version has a narrow screen and poor overview. To be certain, re-check on the web version when you have time.

Note that terminology differs between the mobile and web versions. Topics on the web version correspond to memory files on the mobile version. The functions are equivalent and the storage location is shared. Edits made in either are reflected. Choose whichever is easier to use.

Conditions for Storing Sensitive Topics and

Excluded Information

The recent update clarified the handling of sensitive topics. It is an option called Include sensitive topics in memory. By default, memories concerning health, race, ethnicity, religion, politics, and sexual identity are not saved. Highly sensitive information is recorded only when enabled.

When a sensitive-topic memory is generated, a notification is displayed. Users can grasp on the spot whether it has been recorded. In older versions or mobile apps, the notification may not appear. In that case, the sensitive topic is not saved to memory. In Anthropic’s explanation, this is distinguished as a specification.

“if you’re using an older version or the mobile app, you will not see this notice, and Claude won’t save the sensitive topic to its memory.”

It does not retroactively save past conversations. Only conversations after enabling are covered. Government-issued ID numbers are always excluded. They are not saved even when the option is enabled. It can be evaluated as a design that leaves a safety margin.

Users handling sensitive information must check the settings. Particular caution is required when handling health information for work. If not needed, it is safer to keep the default. If enabling, strengthen monitoring of the topic list. Operation that does not miss notifications is important.

Options to Stop Auto-Generation and Reset

A means to stop automatic generation is also provided. Disable Generate memory from chats. After that, automatic recording from conversations will not occur. It is intended to be used together with manual organization.

After disabling, you can choose Pause memory and Reset memory. Pause is an option to temporarily stop recording. Reset erases accumulated memory. Anthropic warns that the latter cannot be undone.

“Upon re-enabling the feature, you’ll start from scratch, and Claude will not have its previous memory,”

After re-enabling, accumulation starts from scratch. Previous memories are not carried over. It can be used to reset when there is a lot of misinformation. On the other hand, useful background is also lost. Before executing, it is advisable to write out the subjects and preserve them.

For development use, behavior changes depending on whether memory is present. For bug investigation, reproduction conditions must be aligned. In situations verifying tool behavior, as in sqlite-utils 4.0rc2, Claude Fable Finds Critical Bug, stating the memory state helps reproducibility. Stopping or resetting is worth including in verification procedures.

Memory Inheritance in Collaborative Work and

Operational Notes

Memory is carried over between chats and Claude Cowork. The context of document creation and conversation is less likely to be fragmented. It is useful for sharing premises in long-term projects. Conversely, incorrect premises are also carried over. The expanded handover scope increases the importance of inspection.

In operation, be careful about renaming topic names. The same content may be duplicated under different names. Duplication causes inconsistency in responses. Delete the older one and unify into the newer one. Regular deduplication is effective.

On shared devices, consider stopping automatic generation. Conversations with others may mix into memory. Use Pause and resume after work. For highly sensitive projects, Reset is also an option. Appropriate use according to usage pattern is required.

Editorial Opinion

In the short term, we expect memory inventory to become established in practical work. With topic-based editing tools in place, correcting false premises has become easier. Over the next 3 to 6 months, inspection items will be added to manuals for development and support work. To ensure reproducibility of responses, recording of memory status will spread. In tool-integrated settings, we assess that checking settings will become standard practice at the start of work.

In the long term, we expect memory portability and the right to erasure to become issues. On a 1- to 3-year horizon, migration between providers and audit compliance will be required. Users will retain accurate memories as assets and request deletion of unnecessary memories. In design, clear indication of storage scope and means to export by topic will become important. Memory transparency may become established as a criterion for product selection.

The remaining question is who defines the boundaries of sensitive topics. Categories such as health and politics are perceived differently across countries and cultures. Is the current situation, where notification display conditions differ by version and device, sufficient? Will disclosure of the basis for exclusion rules progress? The development of explanations that allow users to choose with confidence is being questioned.

References

• “How to edit Claude’s memory”, by staff@engadget.com (Igor Bonifacic) — Engadget, 2026-09-04T12:00:00.000Z (ARR)
• Source URL: https://www.engadget.com/2246397/how-to-edit-claude-memory/