Anthropic has connected Claude's memory feature across Chat and Claude Cowork. Context built up in a chat conversation now carries over into Cowork tasks, and the reverse holds too. Anthropic describes the design this way: "wherever you work with Claude, it starts from what it already knows about you" — no matter which entry point you use, Claude starts from what it already knows.
The feature is on by default for Free, Pro, and Max subscribers, covering web, desktop, and mobile clients — mobile users need to update to the latest version. Enterprise and Team plans work the other way around: administrators decide whether to enable it, and then individual users can turn it on for themselves.
Memory Is Now Broken Into Editable Entries
The most concrete part of this update is how memory is presented. Users can see everything Claude has saved in the Memory settings, grouped by topic, with each entry readable, editable, or deletable on its own. Anthropic notes that these memory files are deliberately kept short.
Being visible and editable is what separates this from most similar features. Most long-term memory systems on the market rely on an internal summary users never see — what the model has remembered about you, how it has categorized you, with no corresponding interface. When a memory gets something wrong — say, mistaking a one-off technical research task for a long-term interest — users typically have no option but to wipe everything and start over.
Breaking memory into individual entries makes correction more precise: delete the one entry that got it wrong, and keep the rest.
Five Categories Excluded by Default
The privacy boundary here is drawn fairly clearly. Health status, race, ethnicity, religious belief, and political views are five categories Claude does not write to memory by default. Users can actively turn on "include sensitive topics in memory," after which they'll get a prompt each time this kind of information would be saved.
One category is unaffected by the toggle either way: Social Security numbers, government-issued ID numbers, criminal records, immigration status, and content that violates the usage policy are never remembered, regardless of the toggle's position.
This layer of design matters more in practice on the enterprise side. Handing the toggle to administrators on Team plans puts a question on the table that's usually left implicit — who owns the long-term context between employees and AI. In the past this was typically treated as a personal decision, and only got scrutinized after a compliance problem surfaced.
Has the Old Sycophancy Problem Gotten Better?
Memory features have always carried a side effect. Writer's earlier research found that models with memory turned on tend to agree with users more readily: once a model remembers your position, it pushes back less in the next round. The more accurate the memory, the more stable this bias can become.
Editable memory entries aren't a cure, but they're at least an entry point. When you notice Claude getting less willing to push back on a given topic, you can go check what's stored in memory about that topic and delete the relevant entries before trying again. That kind of troubleshooting wasn't possible before, because there was nothing to see.
For people running long tasks in Cowork every day, the repetition this update saves is real — project background, coding conventions, and writing tone only need to be said once. The cost is having to check back occasionally on what's accumulated in memory.
Sources: Anthropic official blog, CocoLoop, Writer research on memory and sycophancy; subscription tiers, platform coverage, and the sensitive-information list should be checked item by item against official announcements.