Anthropic Merges Claude Chat and Cowork Memory Into a Single Shared System

Anthropic announced on August 25, 2026, that it has merged the memory systems used by Claude's chat interface and Claude Cowork, its agentic work product. Information Claude learns in one surface now carries over to the other automatically, with no manual sync required.

What changed

Previously, Claude chat and Claude Cowork maintained separate memory stores. A user's preferences, project details, or organizational context shared in chat stayed locked to chat; anything learned in Cowork stayed locked to Cowork. Under the new system, the two are unified: according to Anthropic, "the memory you use in chat is the same as in Claude Cowork."

Anthropic's own examples illustrate the intended use case. If a user explains in chat how their team defines a metric, every Cowork-built quarterly report afterward uses that definition without re-briefing. If a user brainstorms a conference agenda in chat, Cowork can later build a budget document using the headcount, city, and speaker list already discussed.

Anthropic also changed how memory is written. Claude now adds information to memory dynamically as a conversation happens, rather than summarizing and storing memory only after a session ends.

Claude Code was not part of this update. When asked directly, Anthropic told ZDNET the change is "focused on Claude Cowork and chat" only.

Defaults and privacy controls

Memory is on by default across Free, Pro, and Max plans, on web, desktop, and mobile. Enterprise and Teams accounts have memory off by default, subject to eligibility.

Sensitive-topic storage is handled separately and is off by default for all tiers. By default, Claude does not store information related to health, race, ethnicity, religion, politics, or gender identity. Users can opt in to let Claude retain sensitive information — Anthropic's stated example is a gluten allergy relevant to recipe suggestions — and Claude will display a notice when it does so. Certain categories remain excluded even if a user opts in: government ID numbers, criminal history, immigration status, and anything violating Anthropic's acceptable use policy.

Users can pause or fully reset memory at any time. Settings include a toggle for "Generate memory from chats" and a separate toggle for "Include sensitive topics in memory," both found in a new Memory section of account settings. Saved memories appear as editable text files under a Topics list; editing one — for example, updating a company name — propagates the change to future conversations.

This is distinct from Anthropic's separate "Help Improve Claude" data-training setting, which governs whether conversations are used to train future models and can be disabled under Settings > Privacy for non-enterprise accounts.

What this means

This is a product and settings change, not a new model release — no new weights, checkpoint, or version string is involved. The practical effect is that Claude's cross-session context, previously fragmented by product surface, now behaves more like a single persistent assistant. For users running multi-step work in Cowork after establishing context in chat, this removes repetitive re-briefing.

The tradeoff is expanded default data retention. Memory is opt-out rather than opt-in for individual users, meaning most people will have cross-product memory active without taking any action. Anthropic's sensitive-topic exclusions and notice system provide guardrails, but the core shift — a chatbot that silently accumulates a persistent profile across both conversational and agentic use — raises the same tension every major AI vendor is now navigating: context improves output quality, but it also means more personal information sitting in a system whose retention and training policies most users have never fully audited.