Claude Memory: What It Remembers, and How to Control It
Claude's memory lets it carry what it has learned from one conversation into the next, instead of meeting you as a stranger every time. Turn it on in Settings, and it starts remembering the context you would otherwise re-explain constantly — what your company does, how your team writes, the decisions you already made.
It is inspectable and deletable, incognito chats are excluded, and you can switch it off entirely. There are two kinds worth knowing apart: personal memory, which follows you across your own chats, and agent memory, which lets automations built by your company accumulate institutional knowledge and share it between them.
The re-explaining tax
Without memory, every conversation starts at zero. You explain what your company does. You explain who the client is. You paste in the tone guide, again. You re-establish the decision you already made last Tuesday, because as far as Claude is concerned last Tuesday never happened.
Most people respond to this by writing longer and longer prompts, bolting on another “remember that we always…” each time something goes wrong. It works, sort of, and it is exhausting — and it quietly caps how much anyone is willing to use the tool. Memory removes that tax. You explain something once, and it stays explained.
The two kinds of memory
These get conflated constantly, including by vendors, and they behave differently enough that it is worth ten seconds to separate them.
| Personal memory | Agent memory | |
|---|---|---|
| Who it serves | You, across your own conversations | An automation your company built, across every run |
| What it holds | Your preferences, context, ongoing work | Edge cases, corrections, patterns from past runs |
| Who can see it | You | Whoever administers your Claude setup |
| Shared? | No — it is yours | Yes — agents can share what they learn with each other |
| Turn it on | Settings, one toggle | Configured per agent when it is built |
| Why you care | Stop re-explaining yourself | Corrections stop evaporating and start compounding |
Personal memory is the one nearly everyone means when they search for this, and it is a two-minute change. Agent memory is the one that changes what a company can build.
What it stores, and how to control it
The thing that makes memory workable rather than unsettling is that it is not a black box. It stores what it learned as something you can read, and everything below follows from that.
- You can read what it remembers. If Claude starts behaving differently, you can look at what it picked up rather than guessing.
- You can correct it. If it generalised too hard from one off-hand comment, edit the memory directly instead of arguing with it across three prompts.
- You can delete it. Individual memories or all of it, the same way you would delete a file.
- You can switch it off. Memory is a setting, not a condition of using Claude.
- Incognito chats are excluded. If you want a conversation that leaves no trace, that is what incognito is for.
People turn memory on, mention something once in passing, and are then puzzled when Claude treats it as a standing rule. Memory does not distinguish between “this matters permanently” and “this was true on Thursday.” If something is temporary, say so. If it has become wrong, delete it. Treat it like a shared document your colleague reads before every meeting, because that is functionally what it is.
Agent memory: where it gets interesting
If your company has built automations — something processing tickets, invoices, applications, leads — you have almost certainly hit this wall. The first run is impressive. The tenth run has you spending more time correcting it than the work would have taken, because every correction has to be made again.
The agent does not remember that you said never to CC legal on routine vendor mail. It does not remember that your largest customer prefers Friday updates. It does not remember the exception you carved out for the German subsidiary's VAT. The thing holding agents back was never intelligence. It was amnesia.
With memory, the first correction sticks. Next time the situation arises, the rule gets applied without you. Over months, the agent accumulates the institutional knowledge that currently lives in three people's heads — in a place where it gets applied consistently, and where a new colleague can read it.
And agents can share. The one handling inbound enquiries can pass what it learned about your customers to the one running outbound. The one processing invoices can share what it knows about vendors with procurement. You stop maintaining isolated pockets of knowledge in separate bots.
For regulated work this auditability is what makes the whole thing usable. The AI just knows does not survive a compliance review. Here is the file it consulted, here is when it was written, here is who reviewed it does.
The privacy question, answered properly
Someone on your team will ask, and they should. The honest answer has three parts.
Memory is opt-in and reversible. It is off until you turn it on, and turning it off is one toggle. Nothing about using Claude requires it.
What it holds is visible to you. This is the part that separates it from the thing people fear. You are not trusting a system that has silently formed opinions; you are reading a set of notes and editing the ones that are wrong.
Your account type still governs everything else. Memory does not change the underlying data rules — who can see your conversations, whether they train anything, how long they are kept. Those depend on which kind of account you opened Claude through, and that distinction matters more than memory does. It is the single most misunderstood thing about using AI at work, and worth settling before you worry about this feature.
What to do this week
1. Turn personal memory on and give it the context you retype most
Open Settings, enable memory, then spend five minutes telling Claude the things you explain in every third conversation: what the company does, who the customers are, how you like things written, what you are working on. That five minutes is the highest-return thing in this article and you never repeat it.
2. Read it back a fortnight later
Open what it has stored and check it. You will find one or two things it has overweighted — a passing remark treated as policy. Delete those. This habit, twice a year, is all the maintenance it needs.
3. If you run agents, pick the one you correct most
Find the automation where you keep adding “and also remember to…” to the instructions, or where the same edge case keeps catching it out. That is the one where memory pays off first. Enable it there, work with it for a fortnight, and watch the volume of corrections decay.
What memory does not fix: if an agent has the wrong scope, or is making decisions a human should sign off on, remembering things better will not help. The work of deciding what an agent should do, what it must escalate, and what gets reviewed still belongs to people. Memory makes a well-designed agent noticeably better. It does not rescue a badly designed one.
Working out which of your recurring tasks are worth handing over — and what the agent must never decide alone — is most of what a Deployed Kickstart half-day is for. The Partner programme keeps that work compounding as capabilities like this keep landing.
Frequently asked questions
What is Claude memory?
A feature that lets Claude carry what it has learned from one conversation into the next, so you stop re-explaining the same context every time — what your company does, who the customers are, how your team writes. It is off until you turn it on in Settings, and it can be switched back off at any point.
How do I turn Claude memory on?
Open Settings and enable memory. Then spend five minutes telling Claude the things you find yourself explaining in every third conversation. That five minutes is the highest-return part of setting it up, and you never have to repeat it.
Can I see and delete what Claude remembers?
Yes. What it stores is readable rather than hidden, so if Claude starts behaving differently you can look at what it picked up instead of guessing. You can edit a memory if it generalised too hard from an off-hand comment, delete individual memories or all of them, and switch the feature off entirely.
Does Claude remember incognito conversations?
No. Incognito chats are excluded from memory. If you want a conversation that leaves no trace, that is what incognito is for.
What is the difference between personal memory and agent memory?
Personal memory serves you across your own conversations and holds your preferences and ongoing context. Agent memory belongs to an automation your company built, holds edge cases and corrections learned across every run, is visible to whoever administers your Claude setup, and can be shared between agents so one automation’s learning benefits the others.
Why does memory matter for AI agents?
Because the thing holding agents back was never intelligence — it was amnesia. Without memory, every correction has to be made again on the next run, so by the tenth run you spend more time correcting than the work would have taken. With memory the first correction sticks, and over months the agent accumulates institutional knowledge that currently lives in a few people’s heads.
Is Claude memory safe to use with sensitive information?
Memory is opt-in, reversible, and what it holds is visible to you — so you are reading and editing a set of notes rather than trusting a system that has silently formed opinions. It does not change the underlying data rules, though: who can see your conversations, whether they train anything, and how long they are kept all depend on which kind of account you use, and that distinction matters more than this feature does.
What is the most common mistake with Claude memory?
Mentioning something once in passing and being surprised when Claude treats it as a standing rule. Memory does not distinguish between “this matters permanently” and “this was true on Thursday.” Say when something is temporary, delete it when it becomes wrong, and read back what it has stored every few months.
Found this useful? Send it to someone who needs it.