How to audit what Claude remembers about you (and what it can't show you)
Since March 2026, every Claude tier (Free included) keeps a persistent memory of you across conversations. Convenient. It also means Claude is building a running record of your role, your preferences, and anything you've told it to remember. If you want to know what a product has stored about you, here's how to check.
Quick answer > To see what Claude remembers: open Settings > Memory, where everything is listed under Topics and you can read, edit, or delete each one. Since August 25, 2026 that same memory follows you into Cowork tasks running in the cloud. What you still can't see: which chat produced a topic, or how it changed. Anthropic's own documentation states that "Individual member memory edits aren't logged."
When did Claude start remembering you?
Anthropic rolled memory out in stages. Team and Enterprise plans got it in September 2025. Max got it on October 23, 2025, with Pro following over the next two weeks. Free got it on March 2, 2026. By mid-2026 it's on by default for Free, Pro and Max on claude.ai, the Claude desktop app, and Claude mobile; on Team and Enterprise an owner has to make it available, and then it is off for each member until they turn it on. It does not apply to the API, and Claude Code keeps its own separate, local, file-based memory rather than this one.
It was announced as a feature. It's also a shift in how much Anthropic stores about each user. Memory before this existed only inside individual conversations. Now it persists across all of them.
What exactly does Claude store?
This has changed twice in seven weeks, and both changes matter for auditing. On July 10, 2026, Anthropic replaced the once-a-day synthesized summary with categorized entries updated as you chat. On August 25, 2026 it changed again: Claude now "saves memory as a set of individual topics as you chat, rather than summarizing conversations after they end," and the announcement calls them "a list of files under Topics" in Memory settings.
The migration is finished for individual plans; Anthropic says a small number of Team and Enterprise organizations remain on the legacy summary. Anthropic's help page now states that it has "migrated users off the legacy experience," and it attached a window to the fallback: "until September 9, 2026, you will see the option to export your legacy memory." That window has now closed. The recovery it described was manual, not a restore: you exported the old memory, pasted it back into Claude, and highlighted whatever appeared to have been dropped. If you had months of accumulated context on the old model and did not take the export, there is no documented recovery path now.
None of it is a transcript. You can also tell Claude directly: "remember that I work in regulatory compliance."
Two things the August update added that change the audit question
The first is reach. Memory is now shared between chat and Claude Cowork when Cowork runs in the cloud, so what you mentioned in a chat is in scope for an autonomous task you are not watching. Anthropic is specific about the boundary: "Cowork sessions that run locally on your computer don't use memory."
The second is a filter. By default Claude "does not store topics related to personal or sensitive subject matter, like your health, race, ethnicity, religious beliefs, politics, gender identity, and other similar areas." You can opt in by turning on "Include sensitive topics in memory" in Settings > Memory, and it applies forward only: "Claude saves sensitive topics going forward. Anything from before you turned it on isn't saved retroactively." Turning the setting back off removes the sensitive items already saved. A short list is never stored either way, including government ID numbers, criminal history, financial account numbers, and immigration status. This is a real privacy improvement, and it is worth being precise about what kind: it is a policy filter applied at write time, not a record. You cannot see what it declined to save, or when.
Memory is viewable and editable in Settings > Memory. You select a topic to read it, edit it with the edit icon, or delete it. You can also pause the whole feature (which keeps the existing memory but stops Claude from using or adding to it), or reset it entirely, which permanently deletes everything including project memories. For a single conversation you can also turn memory off before the first message ("+" in the message box, then Memory), and nothing from that chat is saved. One deletion behavior to know before you rely on it: memory is decoupled from the source chats, and Anthropic's docs state that deleting a conversation does not remove the memories generated from it. I unpacked what that changes in a separate piece.
What can't you audit in Claude's memory?
Claude shows you the current state, not the underlying record. You can't:
- See which chat produced which fact. There's no provenance. A topic says "works in regulatory compliance" but doesn't tell you which conversation, on which date, Claude inferred that from. (Claude does show citations back to source chats when it searches your past conversations. That is a citation at the moment of reference, not a source recorded on the stored memory.)
- See that a memory changed. The Topics list shows the current text of each file. Editing one replaces it, and Anthropic's documentation is explicit that "Individual member memory edits aren't logged." An organization's audit log records when an owner turns memory on or off, not what any individual memory said last month.
- Export it as a structured record. Anthropic ships an experimental memory import/export for transferring memory between Claude and other AI services, and memory is included in standard conversation exports. Both move the memory's text, not a fact ledger with timestamps, provenance, or history.
- Hand it to a DPO. A compliance officer can't review the full memory store on behalf of an organization. Only the individual user sees their own.
- Self-host it. The memory lives on Anthropic's infrastructure, with no option to run it yourself.
If you're in a regulated profession, this matters. A lawyer with client matter facts or a consultant under an NDA is now storing fragments of that with Anthropic, and can't fully see what's been kept. "Claude remembers things about me" is fine. "Claude remembers things about me and I can't fully inspect, export, or self-host the record" is a different thing.
Two details in the Team and Enterprise documentation are worth knowing before you plan around this. Memory "isn't available to organizations with HIPAA, public-sector, or custom data retention agreements," so for some of the most regulated buyers the feature is simply off the table. And when an owner turns memory off for the organization, "all existing memory entries for all users are deleted immediately" and permanently. That is a clean privacy default and a sharp edge at the same time: the same switch that protects the organization destroys every user's accumulated context with no export step in between.
How is auditable memory different?
An auditable memory store answers four questions that Claude's Topics list cannot:
- What is remembered? Each fact, individually, in plain language.
- When was it learned? A timestamp on every fact.
- Where did it come from? Provenance: which conversation, document, or explicit instruction it came from.
- Is it still true? A temporal validity range, so out-of-date facts are distinguished from current ones.
This is the bi-temporal model: every fact is a subject/predicate/object triple with a validity window. "Daniel | has_role | compliance officer | valid 2024-01-01 to present." When the role changes, the old fact's validity closes and a new one opens. The history is preserved.
Calmara is one example of a memory store built this way. The full record is exportable as schema-versioned JSON (Article 20 portability, available on every tier including Free), self-hostable via a Docker bundle, and readable in plain English by anyone you choose to share it with, including a DPO. There are a few other systems in this space. The distinguishing question is always the same: can you read every fact, see when it was learned, and take it with you when you leave?
Why this matters even if you're not in a regulated industry
You don't need a compliance officer to want to know what an AI remembers about you. Memory systems shape what the AI tells you next. A stale fact ("lives in Stockholm" when you've moved, "uses Windows" when you've switched) degrades every future answer. The audit isn't paranoia. It's hygiene. The same rule applies if you're building an AI second brain: you can only trust what it tells you if you can see what it kept.
The test for any AI memory system is simple: show me the record, let me edit it, let me take it with me. Claude passes the first two, and two rebuilds in seven weeks have made it better at both. It half-passes the third, and neither rebuild touched the part that would let you prove anything: when a memory arrived, where it came from, and what it replaced. That's worth knowing before you tell it anything you wouldn't want summarized by someone else.
FAQ
How do I see what Claude remembers about me?
Open Settings > Memory on claude.ai, the desktop app, or mobile. Everything Claude remembers is listed under Topics: select one to read it, use the edit icon to change it, or select Delete to remove it. You can also pause memory (it keeps what it has but stops using or adding to it) or reset it, which permanently deletes everything including project memories. To keep one conversation out of memory, start a new chat, click "+" in the message box and turn Memory off before your first message.
Can I export Claude's memory?
Partly. Anthropic offers an experimental memory import/export that transfers your memory between Claude and other AI services (the import is a paste box, so it works from ChatGPT, Gemini, Grok or anything else that can list its memories; imports are documented for Free, Pro, Max and Team on web and desktop), and as of August 2026 memory entries are also included in standard conversation history exports. What moves is the memory content. There is still no structured fact record with timestamps or provenance. Separately, if you were on the legacy memory model, the one-off legacy export in Settings > Memory was available until September 9, 2026; that window has closed.
Does Claude's memory get used to train models?
That's controlled by a setting, and it's easy to get wrong. Under Anthropic's August 2025 consumer terms, whether your Free, Pro, or Max chats are used to train Claude depends on a data-use choice: allow it and your conversations are retained for five years for training; decline and it's the older 30-day retention with no training. Commercial plans (Team, Enterprise, Education) are excluded entirely. Memory is built from those chats, and Anthropic doesn't exclude it separately, so it follows the same setting. Its own documentation says memory is retained "in accordance with existing chat data retention policies." If keeping conversations out of any training pipeline is the priority, the cleaner posture is a memory store that never leaves infrastructure you control.
Is Claude's memory the same as ChatGPT's memory?
They used to be similar. Since OpenAI's June 2026 Dreaming update, ChatGPT's memory continuously rewrites itself in the background, and the saved-memories list is no longer the default (OpenAI still lets you switch back to it under Settings → Memory); the audit situation there is now worse than Claude's. See what ChatGPT's self-rewriting memory costs you for the full comparison.
What does "bi-temporal memory" mean?
Bi-temporal means every fact has two time dimensions: when the fact was recorded in the system, and the real-world time the fact is true. "Worked at Acme 2020–2023" stays in the record after you leave Acme, marked as no longer current. This preserves history rather than overwriting it.
Can my DPO audit Claude's memory on my behalf?
No. Claude's memory is viewable only by the individual user, not by an organization's data protection officer. Anthropic's Team and Enterprise documentation states that owners "can't view or edit a user's individual memories." If organizational audit is a requirement, you need a memory system designed for organizational review, typically a self-hosted one.
Does Claude Cowork see what I told it in a chat?
Yes, if Cowork is running in the cloud. Since August 25, 2026 chat and Cowork share one memory, so context from your conversations is available to a Cowork task and what surfaces in a task carries back to chat. Anthropic states the exception plainly: Cowork sessions running locally on your computer don't use memory.
Does Claude remember my health, religion, or politics?
Not by default. Since August 25, 2026 Claude excludes personal and sensitive subject matter (health, race, ethnicity, religious beliefs, politics, gender identity) from memory unless you turn on "Include sensitive topics in memory." It applies going forward only, nothing from before the switch is added retroactively, and turning it off again removes the sensitive items already saved. Government ID numbers, criminal history, financial account numbers, and immigration status are never saved either way.
What happens to my old Claude memory after the migration?
Anthropic says users have been migrated off the legacy memory experience. If something appeared to have been lost, Settings > Memory offered a one-off export of your legacy memory until September 9, 2026; you pasted it back into Claude yourself, highlighting whatever was dropped. That was a manual reconstruction, not a restore, and the window has closed.
Written by Dan Hagen