Regular use of conversational AI tends to expose a simple friction.
A person explains a project, a colleague's preferences, or a standing constraint in one session, then finds that the next session or a different mode of the same product starts without that background.
The model can still answer questions, but the user supplies the same facts again.
Memory features were added to reduce that cycle.
The idea is that, the features store selected details across conversations rather than treating every thread as isolated. However, different products have handled visibility, update timing, and product boundaries in different ways.
Now, Anthropic aligned two of those boundaries inside Claude. The memory used in ordinary chat is now the same store available to Claude Cowork when Cowork runs a task in the cloud.
In other words, "Claude's memory works everywhere" in Claude chat and Cowork.
Cowork is the company's agentic interface for multi-step knowledge work: users describe an outcome and the system plans and executes across files, connected tools, or the computer itself, rather than answering one prompt at a time.
Context built in chat, such as a manager's preferred update format or details of a conference discussed earlier, can be present when a Cowork task begins.
Information that arises during a Cowork session can flow back into later chats.
Local Cowork sessions do not use this shared store. Claude Code remains on a separate memory system. There is no setting that keeps chat and Cowork memories apart; they now use one system.
The way memories are written has also changed.
Previously the system summarized a conversation after it ended. It now adds topics while the conversation is still underway.
A deadline that moves to September can be available in the next exchange without an explicit instruction to remember it.
Users can still say "remember this" for a specific fact, and everything retained appears in Settings under Memory as a list of topics, each with a short description and an update date. Individual topics can be read, edited, or deleted.
Toggles control whether Claude searches past chats for relevant details and whether it generates new memory from conversations. Memory can be paused so existing entries remain but no new ones are created, or generation can be turned off. Incognito chats stay outside the store.
It's worth noting that since users can use Claude to process sensitive information, the memory-sharing feature doesn't store certain categories by default.
For example, information users provided Claude that concerns health, race, ethnicity, religious beliefs, politics, and gender identity, are not stored.
A separate setting lets a user include them going forward; earlier mentions are not added retroactively, and a notice appears each time such a topic is saved.
Even with that setting on, government identification numbers, Social Security numbers, criminal history, immigration status, and material that violates the acceptable use policy are not stored.
Memory itself is on by default for Free, Pro, and Max plans on web, desktop, and mobile. Team and Enterprise plans leave it off until an administrator enables it. Cowork itself is not part of the Free plan.
The practical result is a single account-level record of topics that both conversational and delegated work can consult, provided the Cowork session is running in the cloud.
Users who want stricter separation can use a different account, pause the feature, rely on incognito chats, or delete individual topics. The change does not alter how Claude Code stores context, and it does not create per-product memory silos inside chat and Cowork.
What the system keeps is now listed in one place and subject to the same editing controls wherever those two surfaces share the store.
