Claude and memory: what it keeps across chats, and how to control it
Ask most people what "AI memory" means and you get one idea. Claude actually has three separate things that all get called memory, and they behave very…
Claude guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on claude.ai
Ask most people what "AI memory" means and you get one idea. Claude actually has three separate things that all get called memory, and they behave very differently. Knowing which is which is the whole game — it decides what carries over, what quietly resets, and what you can delete.
The three things people call "memory"
1. Context window (memory within one chat)
Inside a single conversation, Claude can hold a lot at once — up to roughly 1 million tokens on some models (a token is about three-quarters of a word). That's why a long thread stays coherent. But it is temporary: close the chat, start a new one, and that in-chat context is gone unless something below carries it.
2. The Memory feature (memory across chats)
Separately, Claude can remember details about you between conversations — preferences, ongoing work, how you like answers formatted — so you don't repeat yourself. As of July 2026 this has been rolling out across plans and surfaces rather than arriving everywhere at once, so whether you have it, and exactly how it looks, depends on your plan and app version. Confirm the current state at claude.ai. Treat the specifics below as the shape of it, not a guarantee.
3. Projects (deliberate, scoped memory)
Projects are a persistent workspace: you add documents, instructions, and context once, and every chat inside that Project sees them. This is memory you control on purpose — nothing leaks in by accident, and nothing carries out to unrelated chats.
How they compare
| Context window | Memory feature | Projects | |
|---|---|---|---|
| Scope | One conversation | Your whole account | One Project's chats |
| Lasts after you close the chat? | No | Yes | Yes |
| You choose what goes in | Implicitly (what you type) | Partly automatic | Fully deliberate |
| Easy to view and edit | N/A | Yes — check settings | Yes |
| Best for | Deep single tasks | "Stop making me repeat myself" | Recurring, document-heavy work |
What Claude does not remember by default
This is the honest part, because the failure modes here surprise people.
- A fresh chat is a blank slate unless the Memory feature is on. Claude Code sessions and API calls also start with no memory of past runs unless you wire that in yourself.
- Memory is not perfect recall. It stores summarized notes, not a transcript. It can record something wrong, miss something important, or surface a stale detail — Claude can be confidently wrong about what it "remembers" just as it can about facts.
- It doesn't reason across everything at once. Having a memory stored is not the same as Claude reliably applying it in every reply.
- Different surfaces can differ. What the web app remembers and what the mobile app or the API knows are not guaranteed to be the same pool. Check each.
If continuity actually matters, put it in a Project or paste it into the chat. Don't trust ambient memory to catch it.
The controls you have
You are meant to be able to:
- See what's stored, and edit or delete individual items.
- Turn memory off entirely, or clear it.
- Use a temporary chat for a one-off you don't want remembered — useful for a sensitive question or a throwaway experiment.
Exact menu names and locations change, so check the current settings on claude.ai or in the app rather than trusting a screenshot from months ago.
Two things memory does not do
- Turning off training is not the same as going local. Whether Anthropic uses your chats to improve models is a separate setting from memory, and neither one makes your request local. Claude is cloud-only: your text still travels to Anthropic's servers to get an answer. If that's the concern, see local vs cloud.
- Memory is not encryption or privacy by magic. Deleting a memory removes that stored note going forward; treat anything you send as processed in the cloud.
When this is the wrong tool
- You need a guaranteed, auditable record. Memory is summarized and editable by the system, not a ledger. Keep your own notes.
- You want hard privacy. For anything you'd rather never leave your machine, a local model on your own hardware fits a goal that cloud memory can't — cloud-only means cloud-only.
- The context is huge and structured (a whole codebase, a document set). That's a job for Projects, files, or Claude Code — not for hoping the Memory feature absorbed it.
- You're doing a sensitive one-off. Use a temporary chat instead of teaching your account something you'll want to scrub later.
