Contents
Yesterday Anthropic shipped a change that sounds small in a changelog and is not small in practice: Claude’s memory now works the same way whether you are typing in chat or handing a task to Cowork, its agent that runs multistep jobs in the cloud. Tell Claude in chat that your manager hates bullet points that start with a “loud” verb, and the next time Cowork drafts a report for that manager, it already knows. Explain once how your team defines a metric, and every slide deck Cowork builds from then on uses that definition without you repeating yourself.
That sounds like a UX nicety until you think about what it replaces. Every agentic product today has the same unsolved problem: an agent that starts each task from zero context is only as useful as whatever you can cram into a single prompt, and a team that has to re-explain its conventions every time is a team that quietly stops delegating to the agent. Unifying memory across a conversational surface and an autonomous one is Anthropic betting that persistent, portable context is what actually determines whether people trust an agent with real work, not raw model capability.
Three mechanics worth understanding
Memory updates as you talk, not after the conversation ends. Earlier memory implementations across the industry have mostly worked by summarizing a session once it’s over. Claude’s now writes memories in real time, so if you mention mid-chat that a deadline moved to September, your very next conversation, or the Cowork task you kick off five minutes later, already reflects that without you saying “remember this.”
Everything it kept is visible and editable, topic by topic. Settings > Memory lists what Claude has stored as discrete topics rather than an opaque blob. You can open any topic and edit or delete it directly, or just tell Claude in conversation to change it. That auditability matters more once the same memory is feeding an agent that acts on your behalf in Cowork rather than just chatting back at you.
Sensitive topics are off by default, and some things are never stored regardless. Health, race, ethnicity, gender identity, religion, and political beliefs are excluded from memory unless you flip a toggle in Memory settings, and Claude notifies you the first time it saves something sensitive after you turn that on. Nothing mentioned before you enabled the toggle gets backfilled. A separate, harder line exists underneath that setting: Social Security numbers, government ID numbers, criminal history, and immigration status are never saved, sensitive-topic toggle or not, along with anything that would fall under Anthropic’s Acceptable Use Policy. Worth checking if you manage a workspace: memory is on by default for Free, Pro, and Max plans, but off by default for Team and Enterprise organizations.
Try it: how would a given statement get sorted
The policy above is really a three-way classifier applied to everything you say to Claude. Type a sentence, or pick one of the examples, to see which bucket it falls into under the rules Anthropic described.
About the demo
This runs entirely in your browser using simple keyword heuristics that mirror the categories Anthropic described. It does not call Claude and is not the actual classifier Anthropic uses internally.Notice the middle bucket, sensitive-but-toggleable, is the one that actually requires a decision from you. Everything else the policy handles automatically.
Where I land on this
The unification matters more than the memory feature itself. Plenty of products have “memory.” Fewer have committed to the same memory object being readable by both a chat interface a human drives turn by turn and an autonomous agent that acts unsupervised for minutes or hours at a stretch, because that commitment means every topic you casually mention in chat is now, implicitly, an instruction available to something acting on your behalf later. The topic-level audit view and the real never-store list are the right response to that, but if you run Claude inside a team, the concrete next step is checking your org’s default before someone assumes Cowork already knows something it was never told, and deciding deliberately whether sensitive topics belong in a memory that an agent, not just you, will read.
Source: Anthropic’s announcement and SiliconANGLE’s coverage.