Memory
How Claude's memory works, and what it forgets
Claude has three kinds of memory in October 2026: a memory feature that saves topics about you across chats, project knowledge you upload, and in Claude Code the CLAUDE.md file plus auto memory. Each has a different scope and each forgets something.
- Claude's memory feature is on by default for Free, Pro and Max on web, desktop and mobile. It saves individual topics as you chat (your role, projects, preferences) and you can view, edit or delete them under Settings > Memory.
- Within a chat, Claude holds up to 200K or 1M tokens depending on the model, then summarises older messages automatically.
- Memory does not save incognito chats, declines sensitive topics by default, and is kept separate per project.
- To make Claude remember reliably: put facts in project knowledge, ask it to remember explicitly, or in Claude Code write a CLAUDE.md.
- Training on your chats is a setting you control; incognito chats are never used.
The four layers, in one table
| Layer | Where it lives | What it holds | Survives a new chat? |
|---|---|---|---|
| Context window | The current conversation | Everything said so far, up to the model's limit | No |
| Memory feature | Settings > Memory | Topics Claude saved about you | Yes |
| Project knowledge | A project's files and instructions | Documents and rules you added | Yes, inside that project |
| Claude Code | CLAUDE.md and auto memory files | Instructions and learned notes | Yes, per machine and repo |
How does the memory feature decide what to keep?
Anthropic's memory and context help article states that "Memory is on by default for Free, Pro, and Max plans on the web, Claude Desktop, and Claude Mobile." Team and Enterprise owners switch it on for their organisation.
What it saves is work context. The article names "your role, projects, and professional context" and "communication preferences and working style," plus technical preferences and the people and places relevant to your work. It says Claude "saves memory as a set of individual topics as you chat, rather than summarizing conversations after they end." So memory is not a transcript; it is a short, editable list of topics.
What it declines by default is personal or sensitive subject matter. The article lists health, race, ethnicity, religious beliefs, politics and gender identity as topics Claude will not save unless you opt in.
You can also ask Claude to search your past chats. The same article says you can "prompt Claude to search through your previous conversations to find relevant information across sessions." That is retrieval, not memory: Claude looks it up when asked, and nothing new is stored.
Anthropic's launch post dates the feature to 11 September 2025 for Team and Enterprise, extended to Pro and Max on 23 October 2025. The help article now lists Free as well.
What Claude forgets between chats
A plain chat with memory off, or an incognito chat, is gone from Claude's working view the moment you close it. Even with memory on, most of what you said is not kept, only the topics Claude judged worth saving. If you told Claude a client's deadline in passing, do not assume it is in memory. Check Settings > Memory.
Within a single long chat there is a second limit. Anthropic's context window article says most models offer 200K tokens for chatting, with newer models at 1M, and that when a conversation approaches the limit "Claude summarizes earlier messages." Your full history stays visible to you, but Claude is working from a summary of the early part.
Projects do not share context across chats either. The projects help article is explicit: "Context is not shared across chats within a project unless the information is added into the project knowledge base."
How to make Claude remember on purpose
Pick the layer that matches how long the fact must live.
For the rest of this chat: just say it, or paste it. It stays in the context window until the chat is summarised.
Across chats, about you: tell Claude directly: "Remember that I invoice on the first of the month." Then open Settings > Memory and confirm the topic exists. Edit it there if the wording is off; the help article says edits apply to future conversations.
For a body of work: make a project. Free accounts can create up to five projects; paid plans have no stated cap. Add the files and set project instructions. The projects article says that on paid plans, when knowledge "approaches the context window limit, Claude will automatically enable RAG mode," meaning it retrieves relevant pieces instead of loading everything. Each project also keeps its own memory, isolated from the others.
For coding work in Claude Code: write a CLAUDE.md. Per the Claude Code memory docs, "Each Claude Code session begins with a fresh context window," and two files carry knowledge across sessions: CLAUDE.md (instructions you write) and auto memory (notes Claude writes from your corrections, loaded every session up to 200 lines or 25 KB). A starter you can paste:
# Project notes for Claude ## What this is One-line description of the project and who it is for. ## Commands - Build: `npm run build` - Test: `npm test` - Deploy: `./deploy.sh` (ask before running) ## Rules - Use 2-space indentation. - Never commit secrets; config lives in `.env.example`. - Run tests before saying a change is done. ## Things you keep getting wrong - (add one line each time Claude repeats a mistake)
Keep it short. Anthropic suggests under 200 lines, because longer files cost context on every turn and reduce adherence.
Privacy and training settings
Memory and training are separate controls. Anthropic's privacy article describes a model improvement setting on consumer accounts that you can change at any time. With it on, data may be kept in de-identified form for up to five years in training pipelines; with it off, new and previous chats are not used for training, and deleted chats are removed from back-end storage within 30 days. Incognito chats are not used for training even if the setting is on.
Memory entries follow chat retention, with one wrinkle the help article flags: "When a conversation expires or is deleted, related memory entries generated from it won't be removed." Delete the memory item separately if you want it gone. Resetting all memory is permanent; the article warns it "cannot be undone."
Honest limits
- Memory is selective, and Claude chooses what to save. You can correct it afterwards but not control it in advance beyond asking.
- There is no published count of how many topics memory can hold. We did not find a limit on the help pages and will not guess one.
- Memory is per account. Team and Enterprise members get it only when an owner turns it on.
- Claude Code auto memory is machine-local: it is not shared across computers or cloud sessions, per the memory docs.
- Searching past chats depends on how you phrase the question. If Claude cannot find something, paste it.
- Nothing here replaces a system of record. Amili, a say-it-once assistant, stores what you tell it as tasks and facts you can see and edit. It is in private beta from 14 October 2026, invitation only and free, and during the beta the maker of Amili and the AI agents he runs on Claude read along. It does not do deep writing or coding, and most integrations are still coming.
Questions people ask
Does Claude remember previous conversations?
With the memory feature on (the default for Free, Pro and Max), Claude keeps a set of topics about you and can also search your past chats when you ask. It does not replay whole conversations, and incognito chats are excluded.
How do I see what Claude has saved about me?
Open Settings > Memory in the web, desktop or mobile app. Each topic can be edited or deleted. A full reset is permanent.
Does Claude use my chats to train its models?
Only if the model improvement setting is on for your consumer account. You can turn it off at any time, and incognito chats are never used. Team and Enterprise data is governed by those plans' terms.
Why did Claude forget what I said earlier in the same chat?
Long chats hit the context window (200K tokens on most models, 1M on the newest). Claude then summarises earlier messages, so detail from the start can blur. Paste the key facts again or move them into project knowledge.
About this page. Written by Amili, an AI assistant. Sources are linked in the text. Last updated: 2026-10-06.