Claude Memory: What It Remembers About You and How to Control It

What Claude Memory Actually Stores

Claude memory stores a set of short entries about how you work — your role, your preferences, the projects you keep coming back to — organised into categories you can read. It is not a transcript. Claude decides what is worth keeping from a conversation and writes a summary of it, which means the details it keeps are its interpretation of what mattered, not a word-for-word record.

In practice, entries tend to look like this:

• Your job title, team, and the industry you work in
• Standing format preferences — bullet summaries, no jargon, British spelling
• Tools and systems you mention repeatedly
• Ongoing projects and the people involved in them
• One-off instructions it mistakenly read as permanent rules

That last one is where most of the trouble starts. Ask for a three-bullet version of something on a busy Tuesday and there is a real chance Claude files “prefers three bullets” as a lasting preference rather than a passing request.

The other thing worth understanding is timing. Claude memory is not a filing cabinet you deposit things into deliberately. It is built from ordinary conversation, which means you are adding to it constantly without deciding to. Most people’s memory is therefore an accurate record of what they happened to mention, not what they would have chosen to store.

Where to Find Your Memory Settings

Claude memory settings live under Settings, then Memory. On most personal accounts it now sits as its own menu item, showing a list of individual entries you can open, edit, or delete. Anthropic has been migrating accounts to this newer version, so what you see depends on which one you are on.

If Memory is not a top-level item, look under Capabilities instead. That is the older version, which keeps one rolling summary rather than a list of separate entries. The actions are the same; the screen just looks different.

Team and Enterprise accounts tend to move last, so if your work account looks different from your personal one, that is expected rather than broken. Anthropic’s official help article on memory and chat search documents both versions and is the place to check when the interface shifts under you again.

Meet the Keep-Correct-Contain Method

The Keep-Correct-Contain Method is a three-decision routine for AI memory: Keep only the context that earns a permanent place, Correct entries as they drift out of date, and Contain anything that should never persist at all. It is the approach we teach at PromptPeakAI for any tool that remembers between sessions, and it takes the guesswork out of a settings screen most people find intimidating.

The reason it works is that Claude memory failures are almost never technical. They are editorial. Something got in that should not have, or something true in March is still sitting there in August. No toggle fixes either of those — a decision does.

The first entry is a formatting instruction from one specific task, now applied to everything forever. The second is durable context about the job itself — the kind of background that improves output for months without you touching it.

Most guides stop at “here is the delete button”, which is the least useful part. The delete button is obvious. Knowing which of forty entries deserves deleting is the actual skill, and that is what the three decisions give you.

How to Prune Your Memory

Pruning Claude memory takes about five minutes and only needs doing once a month. The goal is not an empty memory — it is a memory where every entry is still true.

Work through it in this order. Asking first is deliberate: it tells you what you are dealing with before you start clicking things, and it usually surfaces one or two entries you would never have gone looking for.

Before opening any settings, ask directly. It is faster than scrolling a list, and Claude will describe the entries in plain language rather than clipped summaries.

Tell me everything you currently remember about me and my work.

Group it into: my role and organisation, my writing and formatting preferences, my ongoing projects, and anything else.

For each item, say whether it looks like a durable fact or a one-off instruction you saved by mistake. Do not guess — only list what is actually in your memory.

You do not need the settings screen for this. Tell Claude in the conversation and it updates the entry directly, which is far quicker than hunting for the right line in a list.

Update what you remember about me. I have moved from [old role] to [new role], and I no longer work on [old project].

My current priorities are [project one] and [project two].

Also remove anything you saved about formatting preferences from single tasks — those were one-off requests, not standing rules.

For individual entries, open Settings and Memory, select the entry, and either use the box that lets you tell Claude what to change or remove, or delete it outright. If you want a genuinely clean slate, Reset memory wipes everything — including memories held inside your projects — and it cannot be undone.

Global Memory Versus Project Memory

Claude memory comes in two scopes. Global memory applies to every conversation you have; project memory stays inside a single project. Getting this split right is the difference between Claude knowing your job and Claude mixing up two clients in the same sentence.

The rule is simple: global memory holds facts about you, project memory holds facts about the work. Your role, your tone preferences, and your reporting style belong in global. A client’s pricing model, a specific system’s constraints, or a campaign’s approved messaging belong in a Claude project, where they cannot contaminate unrelated chats.

Add the following to this project's context only — do not save it as a general fact about me:

Client: [client name]
Constraints: [approval rules, banned claims, tone requirements]
Key numbers: [figures that appear in every deliverable]

If I ask you about a different client in another chat, none of this should apply.

Keeping Sensitive Work Out of Memory

The safest approach is not to clean sensitive information out of Claude memory afterwards — it is to stop it going in. Incognito chats exist for exactly this: nothing from them contributes to memory, and they stay out of your chat history.

Worth knowing before you rely on that: incognito chats are still covered by your organisation’s data retention policy and can still appear in an organisation’s data export. They keep information out of Claude’s memory. They are not a privacy shield from your employer, which is a distinction the same as the one covered in our guide to whether Claude is safe to use at work.

The other distinction people get wrong is between deleting a chat and deleting a memory. These are two different actions on two different screens. Deleting a conversation is a tidy-up of your sidebar; it is not a reliable way to unwind something you told Claude. If a disclosure genuinely matters, open Settings and Memory and confirm with your own eyes that the entry is gone.

Import Your Memory From Another AI

If you have spent a year teaching another assistant how you work, you do not have to start over. Anthropic’s official memory import page supplies a prompt you paste into your current AI tool, and you then paste its answer into your Claude memory settings.

The process is three steps:

1. Copy the prompt from Anthropic’s import page into the AI tool you are leaving
2. Copy its answer, and read it before you do anything else with it
3. Paste the edited version into Claude’s memory settings

Step two is the one people skip, and it is the one that matters. Whatever comes back is a year of accumulated assumptions, some of which were wrong when they were saved. Importing without reading is how a stale job title ends up in your Claude memory on day one.

If you are coming from ChatGPT, check what you actually had stored there first, using our guide to ChatGPT memory and the separate custom instructions field — the two hold different things, and only one of them transfers cleanly. Once the import lands, run the Keep step from the method above and delete anything that arrived already out of date.

What Claude Memory Cannot Do

Claude memory is a convenience feature, not a knowledge management system, and the difference shows up fast once you start relying on it for anything that matters.

The practical consequence is that memory should never be the only place important context exists. Treat it as a convenience layer over a system you actually control — the argument we make in more depth in our guide to building a personal AI workflow library, and the broader shift toward context engineering as the skill that matters.

On a work account there is one more layer. Organisation owners can control memory settings at the org level, and memory content is included in standard data exports — so the meaningful privacy line is not free versus paid, it is a personal account versus a company-managed one. If you are using a personal Claude account for work tasks, our guide to unapproved AI use at work is worth ten minutes of your time.

Fixing Common Memory Problems

Almost every Claude memory complaint traces back to one of five causes, and all five are fixable from the same screen. Work down the table before assuming anything is broken.

If Claude keeps re-saving something you keep removing, the cause is usually that you are still mentioning it in conversations. Memory is rebuilt from what you say, so the entry comes back until the behaviour changes. This is also where the difference between memory and standing instructions matters, which our guide to system prompts and custom instructions unpacks properly.

Run a memory audit with me.

List everything you remember about me, then flag anything that is: out of date, contradicted by something else you remember, too specific to one task to be a standing rule, or sensitive enough that it should not be stored at all.

Recommend delete, correct, or keep for each one.

Set a recurring reminder for the first Monday of the month. Five minutes of pruning keeps Claude memory working from an accurate picture of your job, and it is the single habit that separates people who find the feature useful from people who quietly switch it off. If you are still deciding whether Claude fits your work at all, our comparison of Claude’s plans and the plain-language introduction to Claude are the right next reads.