Anthropic Unifies Claude's Memory Across Chat and Cowork

Anthropic is merging the memory system shared between Claude chat and Claude Cowork, so that information Claude learns in one surface carries over to the other. The company announced the change on August 25, 2026, alongside a set of controls for visibility, editing, and deletion of stored memory. (TechCrunch)

The memory update also changes how Claude captures context. Rather than summarizing a conversation at its conclusion, Claude now adds individual topics to its memory as the chat unfolds. This incremental approach means the model is building a persistent knowledge store during the interaction itself, not producing a retrospective summary after the fact.

Users can now see exactly what Claude has retained. Anthropic is exposing stored memory contents, allowing users to read, edit, or delete information on any topic the assistant has saved. This addresses a friction point that has been present since persistent context features began appearing in assistant products: the inability to audit or correct what the model has committed to long-term storage.

The privacy posture is layered. By default, Claude will not store personal or sensitive information, including health data, race, ethnicity, religious beliefs, politics, or gender identity. Users who want Claude to remember details in those categories can enable an "include sensitive topics in memory" toggle. Anthropic also states that Claude will never save government-issued IDs, Social Security numbers, criminal history, immigration status, or anything else that violates its acceptable use policy, regardless of user preferences.

Availability is broad. The memory feature is enabled by default across Free, Pro, and Max plans on web, desktop, and mobile. iOS and Android users need to update to the latest app version to access the unified memory.

The unification matters because it removes a context boundary that limited Claude's utility for users working across both surfaces. Someone who provides project details or preferences in a Claude chat session would previously find that context absent when switching to Claude Cowork, and vice versa. The merged memory eliminates that fragmentation, making the assistant function more like a single continuous entity regardless of entry point.

The broader context here is about how companies handle sensitive data in persistent memory systems. The default-off stance on health, political, and identity-related information reflects a conservative posture toward data categories that carry elevated regulatory and reputational risk. The opt-in toggle gives users agency over their own context boundaries, but the hard lines Anthropic has drawn around IDs, Social Security numbers, and criminal history suggest the company has identified specific data types where persistent storage is not worth the compliance exposure, even with explicit consent.

The shift from end-of-conversation summarization to incremental topic capture is a meaningful architectural choice. Summarization-at-close produces a compressed, lossy artifact that is easy to reason about but may drop details the user considers important. Incremental capture during the conversation preserves more granular context but raises a different concern: the user may not know which specific statements the model has selected for retention until they inspect the memory store. The exposed memory view becomes not just a convenience feature but a necessary transparency mechanism.

Looking at the competitive landscape, persistent memory has been an area of active development across assistant products. The tension between context continuity and user control over retained data is not unique to Anthropic, but the granularity of the controls announced here, particularly the per-category sensitivity toggle and the hard exclusions on specific data types, sets a relatively detailed standard for how that tension can be managed.

For developers and teams building workflows around Claude, the unified memory changes the calculus for multi-surface usage patterns. A preference stated in a chat session, a coding convention established in Cowork, or a project constraint mentioned in either context now persists across both. Teams will want to factor this into how they structure interactions, particularly around shared accounts or environments where memory contamination between users could become an issue.

In this author's view, the decision to enable memory by default rather than present it as an opt-in prioritizes continuity of experience over minimal data retention. Users who prefer a clean-slate assistant on each interaction will need to actively manage or disable the feature. Anthropic is betting that the productivity gains from persistent context outweigh the friction of managing stored memory, and for most users that is likely the right call. The audit and edit controls are the safety valve that makes default-on defensible.