Anthropic is giving Claude a more continuous memory across its conversational and agentic experiences.
Since August 25, 2026, information remembered during a standard Claude chat can also be used when the user delegates a task to Claude Cowork. Context learned during a Cowork task can then return to future conversations.
The update addresses a recurring problem with AI agents: users often have to explain the same project, preferences and working relationships every time they move from discussing an idea to executing it.
Anthropic is also making Claude’s memory more transparent. Users can see the information stored about them, correct individual entries, delete specific memories or reset the entire system.
However, the integration has several important limitations. Shared memory only works with Cowork tasks running in the cloud, projects continue to use separate memory spaces, and deleting a conversation does not necessarily remove the memories created from it.
One memory between conversation and execution
Claude Chat and Claude Cowork serve different purposes.
Chat is primarily a conversational interface. It helps users explore ideas, ask questions, analyze information and progressively develop a project.
Cowork is designed for delegation. A user describes an objective, and Claude can coordinate several steps, work with files, use connected tools and return a more complete result.
Until now, moving between those experiences could create an awkward break in continuity. A user might spend several conversations explaining a project before having to provide the same information again when asking Cowork to act on it.
The new memory system is intended to remove some of that friction.
Imagine that a user has discussed an upcoming conference with Claude, including its location, speakers, schedule and communication style. When asked to prepare an agenda in Cowork, Claude can now use the context already developed in Chat instead of requesting another complete briefing.
Information also travels in the other direction. If Cowork learns an important project detail while completing a task, that information can become available in subsequent Claude conversations.
This reflects a wider shift from AI systems that only answer questions to assistants capable of carrying context into practical work. Demystia explored the same transition in its comparison of ChatGPT Chat and ChatGPT Work.
Claude does not simply remember every conversation
Shared memory does not mean that Claude stores a complete copy of everything a user has ever said.
Anthropic distinguishes between two separate capabilities:
- memory;
- searching and referencing previous chats.
Memory extracts selected information that may be useful in future interactions. This can include a user’s role, ongoing projects, communication preferences, technical choices and working style.
Chat search works differently. When asked about an earlier discussion, Claude can search eligible previous conversations and retrieve relevant passages. Anthropic says these searches use retrieval-augmented generation and appear as tool calls inside the conversation.
This distinction matters because memory is available on Free accounts, while searching previous chats is reserved for paid Pro, Max, Team and Enterprise plans.
A remembered preference is therefore not the same as Claude searching an entire conversation history.
Memories are now organized into individual topics
Anthropic has also changed how Claude builds memory.
The previous system created a broader summary after conversations ended. The new version records information as individual topics while the conversation is still happening.
If a project deadline changes, for example, Claude can update the corresponding topic without waiting for a later summary. Users can also explicitly ask Claude to remember an instruction or preference.
These topics can cover information such as:
- a person’s role and professional context;
- active projects and objectives;
- colleagues or other important contacts;
- preferred writing and communication styles;
- programming languages and technical preferences;
- deadlines and other ongoing project details.
The objective is not to preserve a transcript. It is to maintain a practical working context that can improve future responses and Cowork tasks.
Users can see and edit what Claude remembers
Every saved memory is visible under Settings > Memory, where the information is grouped by topic.
Users can open an individual topic to read what Claude has stored. They can then edit incorrect or outdated information or delete the topic entirely.
Memory can also be managed conversationally. A user can tell Claude to remember a new preference, correct an existing detail or forget something. According to Anthropic, the change applies from the next conversation onward.
Two broader controls are available:
- Pause memory prevents Claude from using existing memories or creating new ones, while preserving the information already stored.
- Reset memory permanently deletes all memories, including project memories, and allows the user to start again from an empty system.
The reset operation cannot be reversed.
These controls give users more visibility than an opaque system that silently builds a profile in the background. They also make it easier to correct information that was once accurate but is no longer relevant.
Projects still maintain separate memories
The expression “one memory” requires another qualification.
Each Claude project continues to have its own memory space and project summary. Context associated with one project is kept separate from other projects and from conversations taking place outside a project.
This separation can be useful for people working with several clients, companies or personal activities. A writing preference or confidential detail belonging to one project should not automatically influence an unrelated assignment.
Claude’s memory is therefore unified between Chat and Cowork within the relevant context, but it is not one unrestricted database in which every project is mixed together.
Shared memory only works with Cowork in the cloud
The most important technical limitation concerns where Cowork runs.
Anthropic’s documentation specifies that shared memory is available when Cowork performs a task in the cloud. Cowork sessions running locally on the user’s computer do not use the shared memory system.
This difference could surprise people who move between cloud and local workflows.
Cloud Cowork sessions can access remembered context from Chat and send new information back into memory. A local Cowork session must instead rely on the instructions, files and context explicitly available in that session.
Users dealing with sensitive local files may consider that separation beneficial. However, it also means that Claude will not provide exactly the same continuity across every Cowork environment.
Sensitive topics are excluded by default
Anthropic says Claude does not normally save certain sensitive topics to memory.
The excluded categories include information related to:
- health;
- race or ethnicity;
- religious beliefs;
- political views;
- gender identity;
- similar sensitive personal matters.
Users who want Claude to remember this type of context can activate Include sensitive topics in memory. Sensitive information is only stored going forward and is not added retroactively from older conversations.
When Claude saves a sensitive topic, the interface displays a notification so the user can review the entry or change the setting. Turning the option off removes sensitive memories that were already stored.
Some information is never saved, even when the user explicitly requests it. Anthropic lists government identification numbers, financial account numbers, criminal history and immigration status among the prohibited categories.
Incognito chats provide another layer of separation. Conversations started in incognito mode do not appear in normal chat history and do not contribute to Claude’s memory.
Deleting a conversation may not delete its memories
One of the most important privacy details is easy to overlook.
When a normal conversation expires or is manually deleted, memory entries previously generated from that conversation are not necessarily removed.
A user who wants to eliminate the information completely may therefore need to perform two separate actions:
- delete the original conversation;
- locate and delete the related memory topic.
Anthropic says all memory information follows its existing chat-data retention policies and is included in account data exports.
For Team and Enterprise accounts, organizational retention policies also apply. Incognito conversations do not contribute to memory, but Anthropic notes that they may still appear in organizational exports and remain subject to the organization’s retention rules.
Memory settings depend on the subscription
Memory is enabled by default for individual users on the Free, Pro and Max plans across the Claude website, desktop application and mobile apps. Mobile users may need to install the latest version of the application.
The situation differs for organizations.
On Team and Enterprise plans, memory is disabled by default at the organizational level. An owner must first make the feature available before individual members can activate and manage their own memory.
Anthropic says organization owners cannot read or modify the individual memories of their members. They can, however, disable memory for the entire organization. Doing so permanently deletes the existing memory entries of all users in that organization.
Memory is not available to certain organizations using HIPAA configurations, public-sector environments or custom data-retention agreements.
These restrictions reflect the additional compliance requirements involved when AI systems retain professional context across multiple tasks.
Some users should check their legacy memory before September 9
Anthropic has migrated users from its previous memory system to the new topic-based experience.
The company warns that some information may have been lost during that transition. Users who believe Claude has forgotten important context can temporarily export their legacy memory from Settings > Memory.
This export option is scheduled to remain available until September 9, 2026.
After exporting the older memory, users can paste the relevant information back into Claude and ask it to retain the missing context. Anyone who has used Claude’s memory extensively should therefore review the new Topics section before the deadline.
Why shared memory matters for AI agents
Memory can appear less impressive than a new model or benchmark, but it is becoming a central part of useful AI assistants.
An intelligent system may be capable of solving a difficult problem while still being frustrating to use if it repeatedly forgets the user’s preferences, projects and earlier decisions.
This problem becomes more significant with agents.
A chatbot generally answers a single request. An agent may work across several files, applications and stages of a project. To operate effectively over time, it needs more than reasoning ability: it needs a reliable understanding of the context in which the work is taking place.
Shared memory allows Claude to connect two phases of the same workflow:
- understanding and developing the objective through conversation;
- carrying out the work through Cowork.
That continuity could make the assistant feel less like two separate products and more like one persistent collaborator.
It also increases the importance of user control. The more context an AI remembers, the more useful it can become — but the consequences of an incorrect, outdated or overly sensitive memory also become greater.
Claude is becoming a more continuous assistant
Anthropic’s update does not give Claude unlimited memory, nor does it allow every version of Cowork to retrieve everything a user has previously discussed.
What it does provide is a clearer bridge between conversation and action.
Claude can retain useful context across Chat and cloud-based Cowork tasks, while users can inspect individual memories and decide what should be corrected, removed or excluded entirely.
The system’s effectiveness will depend on how accurately Claude selects what deserves to be remembered. Its privacy will depend just as much on whether users understand the distinction between conversations, remembered topics, projects and organizational retention policies.
With this update, the competition between AI assistants is no longer only about which model produces the best immediate answer. It is increasingly about which assistant can understand a person’s work over time — without taking control of that context away from them.