Claude AI Long-Term Memory: How It Works and Settings

Claude AI long-term memory is a feature of the Claude apps that saves what you tell it as individual memory topics and uses them in later chats. Paid plans can also search past chats on demand. You control both in Settings > Your Claude (Settings > Memory in the older layout), where you can view and edit each topic, pause memory or reset it. Memory is on by default for Free, Pro, Max and Ultra, on web, desktop and mobile.

That covers the chat apps. Developers building on the Claude API get a different tool, and Claude Code has its own memory files. All three are below. Facts come from Anthropic’s help center and docs, checked on 8 October 2026; memory settings and project rules re-checked on 1 November 2026.

Does Claude have long-term memory?

Yes. Claude has two ways to remember across conversations. Memory saves facts and context as topics you can read and edit, and Claude uses them in new chats. Chat search lets Claude look up relevant past conversations with a retrieval tool call when you refer to earlier work. Both are optional and adjustable in Settings.

The details, per Anthropic’s chat search and memory article:

• Memory “saves memory as a set of individual topics as you chat, rather than summarizing conversations after they end.” You can also ask Claude directly to remember something.
• Chat search works across non-project chats, and within a project for that project’s chats. It uses RAG and shows up as a tool call in the conversation. It’s on Pro, Max, Ultra, Team and Enterprise, not Free.
• Projects each have their own memory, but it is no longer fully walled off. Inside a project, Claude uses that project’s memory plus your memory from regular chats (never another project’s). New memory is saved to the project, and Claude can reference it outside projects too. To keep a project fully separate, turn off Use account memory under Memory on the project’s page.
• Cowork shares the same memory as chat, but only when a Cowork task runs in the cloud. Cowork sessions that run locally on your computer don’t use memory.

What changed since early October 2026

Anthropic’s help article shows “Updated this week” as of 1 November 2026. The changes that matter if you followed older guides:

The new memory experience isn’t available to organizations with HIPAA or public-sector agreements.

When memory came to each plan

On Team and Enterprise, owners decide whether memory is available, and it starts off for each member until they turn it on.

What Claude remembers, and what it won’t

Claude’s memory leans toward work. Anthropic’s import guide says Claude “may not retain imported personal details unrelated to work.”

Sensitive topics are off by default. Claude doesn’t store personal or sensitive subjects unless you turn on “Include sensitive topics in memory” in Settings > Your Claude. Then it shows a notice each time it saves one. Some data is never saved even if you ask: government ID numbers, criminal history, financial account numbers and immigration status.

Deleted chats don’t delete memories. When a conversation expires or you delete it, memory entries generated from it stay. Delete those entries yourself in the memory panel. This is a change from the older “legacy” memory, which removed deleted conversations from its synthesis and updated every 24 hours.

How to turn Claude memory on, off or reset it

Steps for the current memory experience on web and desktop:

1. Open Settings > Your Claude (or Settings > Memory in the older Settings layout).
2. Toggle Memory from chats (older layout: Generate memory from chats) to turn memory on or off.
3. Toggle Search and reference chats to control chat search (paid plans).
4. Use the memory panel to read topics, edit them, or tell Claude what to change or remove.
5. To stop memory without deleting it, pause it. Existing memory stays, but nothing new is added and nothing is used.
6. To wipe everything, reset memory. This “permanently deletes all memories including project memories” and can’t be undone.

For a single conversation, two options:

• Incognito chat (the ghost icon) isn’t saved to memory or history, and Claude won’t find it in search later. It’s on every plan.
• Memory off for one chat (the “+” menu, before your first message) keeps the chat in history and searchable but leaves it out of memory.

A few Team and Enterprise organizations still use legacy memory, which lives under Settings > Capabilities instead.

How to import memory into Claude from ChatGPT or Gemini

Anthropic’s import and export guide lists four steps, available on Free, Pro, Max and Team plans on the web and Claude Desktop:

1. Export from the other assistant. Use Anthropic’s suggested prompt, which starts “I’m moving to another service and need to export my data” and asks for one code block of dated entries. Edit it to leave out anything sensitive.
2. Start the import. In Claude, open Settings > Memory and select Start import.
3. Paste and add. Paste the text and click Add to memory. Claude stores it as individual entries.
4. Review. Check the entries in the memory panel, then click See what Claude learned about you for a summary chat.

Limits to know: only memory transfers, not chat history. The feature is experimental, so Claude may not keep every item. To export Claude’s memory, ask in a chat: “Write out your memories of me verbatim, exactly as they appear in your memory,” and save the result. Memory data is also included in a full account data export.

If you’re coming from ChatGPT, the guide on how to find ChatGPT memory shows where to view what it saved before you export.

Claude’s context window in the apps

Memory and the context window are different things. The context window is how much text Claude can read in one conversation; memory is what carries across conversations. The context window guide explains the difference.

Anthropic’s context window article for paid plans (Pro, Max, Ultra, Team, Enterprise; updated 30 October 2026, chat sizes unchanged) lists sizes by model in chat:

Claude Code and Cowork have their own per-model limits. On paid plans with code execution enabled, Claude automatically summarizes earlier messages when a chat nears its limit, which “allows conversations to continue indefinitely in most cases.” The full history stays available to reference, but these long chats use more of your usage limit.

Memory for developers: the Claude API and Claude Code

The app memory above isn’t exposed through the API. Developers have two separate mechanisms.

The memory tool in the Claude API

The memory tool lets Claude create, read, edit and delete files in a /memories directory across conversations. It runs client-side: Claude asks for an operation (view, create, str_replace, insert, delete, rename) and your code executes it against storage you control. It’s available on Claude 4 and later models, with tool type memory_20250818. The Python SDK ships a local-filesystem implementation:

 1import anthropic
 2from anthropic.tools import BetaLocalFilesystemMemoryTool
 3
 4client = anthropic.Anthropic()
 5memory = BetaLocalFilesystemMemoryTool(base_path="./memory")
 6
 7runner = client.beta.messages.tool_runner(
 8    model="claude-opus-5-5",
 9    max_tokens=1024,
10    messages=[{"role": "user",
11               "content": "Remember that Acme Corp prefers email follow-ups."}],
12    tools=[memory],
13)
14print(runner.until_done().content)

If you write your own handler, Anthropic’s docs require rejecting any path outside /memories to block directory traversal. The tool also pairs with compaction for long-running agents: compaction shrinks the conversation, memory keeps what must survive it. In Claude Code, a local proxy such as Headroom can also shrink tool outputs before they reach the model, which delays the point where compaction is needed.

Claude Code memory

Claude Code uses CLAUDE.md files that you write (project, user or organization rules) and auto memory, notes Claude writes itself about your preferences, corrections, project context and references. Auto memory lives per repository in ~/.claude/projects/<project>/memory/, and the first 200 lines or 25KB load at the start of each session. For memory that searches past sessions instead of a capped file, plugins such as Hindsight’s build a per-repo memory bank from git history and earlier sessions and inject the relevant parts as Claude Code starts work; see AI coding agent memory. Code-graph tools such as CodeGraph are a different layer: they index the repo’s structure so Claude Code finds code without grepping, but they don’t remember past sessions. Workflow bundles add their own layer too: Everything Claude Code (ECC) saves session summaries with hooks and can learn small habits (“instincts”) from your tool use, though that learning is off by default.

How Claude’s memory compares

Against other chatbots, Claude’s strengths are editable topics, per-project memory you can isolate, and clear per-chat controls. ChatGPT pulls from more sources, and Gemini’s saved-facts feature is narrower; see best chatbot for memory for the side-by-side and Gemini AI long-term memory for Google’s version.

If you’re building your own assistant on Claude and need memory per end user, the memory tool gives you files, not search or fact extraction. Teams often add a memory layer such as Mem0, Zep, Letta or Hindsight for that. Hindsight’s wrap_anthropic() wraps the Anthropic SDK client so memories are recalled before each call and retained after, though each write makes an extra LLM call; the AI agent memory guide explains the options.