Anthropic Links Claude and Cowork Memories with Safeguards

Claude's memory is no longer compartmentalized: the chat and the Cowork agent now share the same context. Anthropic supports this change with default filter adjustments, categorical exclusions, and administrator control over certain plans. The settings also allow for the suspension, correction, or limitation of what is retained.

Anthropic Prevents Automatic Recording of Sensitive Data

By default, data deemed sensitive — such as health, ethnic origin, religious beliefs, political opinions, and gender identity — is not recorded. Their integration requires specific settings, with Anthropic explaining this choice by the varying significance these pieces of information may hold for different users. When such a recording occurs, an alert is sent, and it only pertains to subsequent interactions. Certain information is never stored, regardless of the settings: this includes sensitive identification numbers, such as official identity documents, criminal records, immigration status, and any content incompatible with Anthropic's usage policy. The unification of memory eliminates the distinction between personal and professional uses of a single account, with control now exercised over recorded topics or through the ability to pause memory.

Activation Based on Plans and Availability on Platforms

The unified memory is activated by default for Free, Pro, and Max plans, on the web, desktop application, and mobile versions. An app update is required on iOS and Android to take advantage of the new feature. On Team and Enterprise plans, memory remains disabled until the administrator activates it.

Shared Context Between Chat and Agent During Interaction

Anthropic has merged the memories of the Claude chat and the Cowork agent. The context accumulated in conversations is accessible as soon as a task is assigned to the agent, and the information produced during task execution feeds into subsequent exchanges. Anthropic states that Cowork now has the same memory as the chat and presents this evolution as a means to limit repetitions. The context built over several months of discussions can thus be utilized by the agent, and the elements displayed in Cowork are automatically reflected in the chat. The update also changes how memory functions: it now builds in real-time throughout the discussion, whereas previously it relied on a summary generated at the end of the session. Therefore, if a project deadline is mentioned during a conversation, it will be taken into account in the next exchange without the need to explicitly request its memorization. This evolution applies only to chat and Cowork, with Claude Code not being affected.

Dedicated Settings, Targeted Corrections, and Pause at Will

All settings are centralized in the Memory tab. A button allows for the creation of memory from discussions, while a second controls the addition of sensitive topics. This same screen features the profile card, response options, and the list of recorded themes, each of which can be deleted separately. An input area is available to specify exactly what should be modified or deleted. The cards are brief and apply to all uses: correcting a name in a topic is sufficient for the correction to be taken into account in subsequent conversations. Memory can be paused or completely reset at any time.