What Claude Actually Remembers: Memory, Projects, and Chat Search

Four Layers, Not One

"Does Claude remember?" has four different answers depending on what you mean. Mixing them up is why people are surprised when Claude recalls a detail they never repeated, or forgets a decision they made yesterday.

How Claude Memory Works

Per Anthropic's help center, memory is on by default for Free, Pro, and Max users. On Team and Enterprise, owners decide whether it is available, and each member starts with it off until they turn it on.

The current memory experience saves individual topics as you chat rather than summarizing conversations after they end. You can also say "remember this" to save something explicitly, or ask Claude to change or forget something. (Some accounts still have the legacy experience, which refreshes a single summary every 24 hours.)

To review it, go to Settings > Your Claude and choose Manage next to "Your memory" (older layouts use Settings > Memory). Memories appear as topics you can open, edit, or delete. The "Generate memory from chats" toggle pauses memory: existing memories stay, but Claude stops using them and stops saving new ones. Resetting memory permanently deletes all memories, including project memories.

Sensitive topics are excluded by default unless you enable "Include sensitive topics in memory," and Anthropic says certain data, such as government ID and financial account numbers, is never saved.

Project Memory and Isolation

Each project has its own memory. Inside a project, Claude uses that project's memory plus your general account memory, but not other projects' memory, and new memories formed in a project are saved to that project.

If you need a project to be fully isolated (client work, a confidential launch, or a test where you want Claude to behave as a fresh assistant), turn off "Use account memory" on the project's page. That stops your personal preferences and unrelated work context from leaking in.

Project memory complements the knowledge base; it does not replace it. Memory captures what Claude noticed. Knowledge files capture what you decided. For anything that must be exactly right, like a price, a date, or a scope boundary, put it in a file. The project knowledge base setup guide covers how.

Chat Search: Referencing Past Conversations

On paid plans (web, desktop, and mobile), Claude can search your previous chats when they are relevant. Searches show up as tool calls in the current chat, and Claude cites the source conversations so you can open them. Outside projects, it searches your non-project chats; inside a project, it searches that project's conversations and does not reach into other projects.

You can turn this off with "Search and reference chats" in Settings > Your Claude. Search is often the most useful layer for "what did we decide about X last month?" questions, because it retrieves the actual conversation instead of a summary. Ask for it directly:

Search my past chats for the pricing discussion about the
annual plan (sometime in September). Quote the final numbers
we settled on and link the chat. If you find conflicting
versions, show both and tell me which came later.

The last sentence guards against the main failure mode: search can surface an earlier draft of a decision that you later reversed.

Controlling What Gets Remembered

• Incognito chat: start one with the ghost icon in the upper right of a new chat (outside a project). It is not saved to history or memory and will not be searched later. Available on all plans. On Team and Enterprise, incognito chats still appear in organizational exports and follow retention policies.
• Memory off for one chat: turn off Memory in the "+" menu before sending the first message. The chat stays in history and remains searchable, unlike incognito.
• Delete a memory topic:Settings > Your Claude > Manage, then delete the topic.
• Pause everything: switch off "Generate memory from chats."
• Isolate a project: turn off "Use account memory" on the project page.

Anthropic also offers an experimental memory import and export for moving memories between assistants.

A Memory Hygiene Checklist

Memory is convenient until it is wrong. An outdated memory ("works at the old company," "prefers bullet points" from one formatting request) quietly shapes every answer. A short monthly review keeps it useful.

1. Open Settings > Your Claude > Manage and skim the topic list.
2. Delete topics that are no longer true: old jobs, finished projects, abandoned tools.
3. Edit topics that are half right instead of deleting them; a precise memory is better than none.
4. Check that nothing is there you would not want an answer to reflect, especially on a shared or work device.
5. For each active project, decide whether it should use account memory or stay isolated.
6. If you rely on chat search, keep chat titles meaningful by renaming important chats, so citations are easier to recognize.

On Team and Enterprise plans, there is an extra consideration: owners control whether memory is available, and incognito chats still appear in organizational exports and follow retention policies. Incognito keeps a chat out of your Claude memory and history; it does not take it outside your company's data governance.

One more practical rule: when a fact must be right, state it in the current chat or project rather than hoping memory surfaces it. Memory is good at "how this person likes to work." It is not a database of numbers and dates, and it was not designed to be one. If you catch Claude using a stale memory, correct it in chat ("forget that I use Jira; we moved to Linear") and confirm the topic changed in settings.

What Claude Does Not Remember

1. Exact wording of old chats, unless chat search retrieves them.
2. Anything from incognito chats.
3. Another project's memory or conversations, from inside a project.
4. Details that never became a memory topic. Memory is selective; it is not a transcript.
5. Early details of very long chats, which may be summarized to make room for new content.

A practical scenario: you spend three weeks on a product launch across 40 chats. Memory will likely know you are working on a launch and prefer concise drafts. Chat search can find the chat where you chose the launch date. Neither guarantees that chat 41 uses the final messaging framework unless it lives in a project file or you point Claude to it. For cross-vendor comparisons, see Personal AI Memory Explained.

Where ClaudeKit Fits

Memory and search are Claude deciding what matters. Sometimes you want to decide. ClaudeKit lets you pin any individual message, yours or Claude's, to the Workspace panel inside claude.ai, so the answer that settled a question is one click away instead of buried in a long thread. When you move to a new chat, the Handoff Brief carries a compact summary forward, and chat export (Markdown, JSON, or plain text) gives you a copy you own. All of it is stored locally in your browser, with no account required.

For keeping key answers findable, read how to pin important Claude responses.