Claude App Memory: Settings, Topics, and Controls

What the app remembers

Claude app memory is built from topics, not transcripts. As you chat, Claude writes short files covering things worth carrying forward: your role, the projects you are working on, professional context, communication and technical preferences, coding style, and ongoing work. Those categories are work-oriented, and they are not a promise that every fact will be retained or used.

The useful mental model is not "Claude has a database of every detail." It is "Claude keeps a small set of notes it can consult in future conversations, and you can read all of them." Since the September 2026 update, the same notes are used by Claude Cowork.

Where to control it

Memory controls live in Settings > Memory. From there you turn memory on, browse the topic list, edit or delete individual topics, choose whether sensitive topics are included, and pause or reset memory. Reset is the destructive option: it permanently deletes memories, including project memories.

You can also tell Claude in chat what to remember or what to change. For durable preferences, make the instruction concrete: "remember that I prefer concise meeting summaries with decisions first" is better than "remember my style." Because topics are editable files, correcting a fact once fixes it everywhere rather than requiring a correction in each new conversation.

Incognito, sensitive topics, and deletion

Incognito chats are the clean-room option. When enabled, Claude will not save the chat to memory or chat history. The Help Center also describes deleted conversations being removed from memory.

Sensitive categories are excluded by default. Claude does not store topics about health, race, ethnicity, religious beliefs, politics, gender identity, or similar areas unless you turn on the sensitive-topics setting, and some categories stay excluded either way. The memory everywhere guide covers the exact boundaries.

For sensitive work, use both product controls and ordinary operational hygiene: confirm memory state, skim the topic list, use incognito where appropriate, and do not paste secrets or regulated data unless your plan and policies explicitly allow it.

What not to store

Memory should hold stable collaboration context, not passwords, access tokens, private customer data, or one-off facts that will expire quickly. If a fact has a short shelf life, store it in the live project artifact instead: an issue, task tracker, runbook, or source-controlled doc.

When memory and the source of truth disagree, the source of truth should win. Memory is a convenience layer, not a records system.