Claude Unifies Memory Across Chat and Cowork

Anthropic has unified Claude’s memory across chat and Cowork, allowing context from web, desktop, and mobile conversations to carry into its cloud-hosted computer-use agent. The August 25, 2026 update also adds topic-level inspection, pause and reset controls, and stricter handling of sensitive information.

For developers building workflows that move from planning to execution, the change removes a persistent handoff problem. Claude can retain organizational structures, manager preferences, client histories, and project metrics from earlier conversations, then use them when Claude Cowork performs tasks in the cloud. Learnings from Cowork flow back into chat memory, creating one persistent context across both surfaces.

Memory Across Chat and Cowork

Dynamic memory is the central behavior. Users no longer need to issue an explicit “remember this” instruction for Claude to update its memory store as conversations develop.

The result is a shared context layer rather than two isolated session histories. A conversation can establish how a team reports metrics, while a later Cowork task applies those preferences during desktop work. When the agent encounters new project details during execution, those details can become available in future chat sessions.

The scope has a clear boundary. The unified system applies to Claude chat on web, desktop, and mobile, plus Claude Cowork. It does not apply to Claude Code, which retains its separate file-based memory approach through CLAUDE.md and local Auto Memory. Developers managing coding workflows should treat Claude Code memory and Claude’s consumer-facing memory as separate systems.

This distinction matters for architecture. A team using Cowork can rely on account-level context across conversational and desktop-agent tasks, while a coding workflow still needs repository or local conventions managed independently. Guidance on adding memory to AI agents remains relevant when you are designing your own runtime rather than relying on Claude’s product-level storage.

User Controls and Data Boundaries

Claude now organizes saved memories by topic under Settings > Memory. Users can inspect, edit, or delete individual facts globally instead of treating memory as an opaque summary.

The Pause Memory control stops both new memory creation and memory retrieval while preserving existing entries. Reset Memory permanently removes all user and project memories across the account. The difference gives users a reversible operating mode for temporary privacy needs and a destructive option for a full cleanup.

Sensitive subjects are excluded by default. Claude avoids saving health details, race, ethnicity, religious beliefs, political views, and gender identity unless the user enables Include sensitive topics in memory in settings or through an in-chat prompt. Claude displays a notice whenever it saves a sensitive-topic entry after the setting is enabled.

A hard blocklist remains active even with sensitive memory enabled. Social Security numbers and other government IDs, criminal history, financial account numbers, and immigration status cannot be stored in memory. This gives agent workflows a defined boundary for personalization, although users still need to avoid placing regulated data into prompts and connected applications unnecessarily.

Availability and Migration

Memory is enabled by default for Free, Pro, and Max plans across web, desktop, iOS, and Android. Enterprise and Team plans keep the feature off by default and require administrator enablement, which places the initial decision with workspace governance rather than individual users.

Anthropic has migrated legacy daily memory summaries to categorized, entry-based memory files. Users who were migrated can export their legacy summaries from Settings > Memory until September 9, 2026.

If you automate desktop work with Claude, Cowork workflow guidance should now account for persistent context crossing the chat-to-agent boundary. Audit which facts the agent is allowed to retain, keep sensitive-topic memory disabled unless the workflow requires it, and document Claude Code memory separately from Cowork memory in your deployment policies.