Large language models do not remember anything. Every new conversation starts from zero. That is the single most important thing to understand before using Cowork seriously — and it is also the thing that most new users get wrong.
Cowork works around this statelessness by auto-generating a memory file about you in the background. It scans your chats, writes down what seems important, and silently injects a summary into every new Cowork prompt. That’s useful, but it means you should know what is being remembered, where it lives, and how to change it.
Why the “intern” analogy breaks down
A common comparison: an LLM is like an intern. You can tell it things once and it will remember. That comparison is wrong.
Imagine you asked an intern to prepare a deck in 14-point Arial. The next time you asked them to make a deck, you’d expect them to remember. An LLM does not. If you start a new chat tomorrow and say “make me a deck,” it will default back to 12-point Times New Roman unless either (a) you put the preference in the prompt, (b) you put it in global instructions, or (c) it happens to be in the auto-generated memory.
That last path — the auto-generated memory — is what this module is about.
Where your memory lives
To see what Cowork knows about you:
- Open Settings
- Click Capabilities
- Find Generate memory from chat history and click to view the file
What you’ll see is a plain-text document with sections like:
- Work context — who you are professionally, what you’re working on
- Personal context — interests, family, location, preferences
- Earlier context — older conversations it thinks still matter
- Longer-term background — your career history, as Cloud has reconstructed it
The file was generated by Claude, not written by you. It will be right about a lot of things and wrong or stale about others. Read it.
What happens when you prompt
Every time you send a Cowork prompt, three things get stitched together behind the scenes:
- Your prompt
- Your global instructions
- A summary pulled from the memory file
Claude sees all of this as one long prompt. This is why Cowork “remembers” you are an investment analyst even when you haven’t said so in the current chat — the memory file mentioned it, and that got injected.
Editing your memory
Click the pencil icon next to any line in the memory file. You can:
- Delete a line
- Rewrite it
- Add a new fact (e.g., “I have three kids”)
- Type a natural-language instruction (“forget that this company was my client”)
Submit the change and the file updates. The next Cowork prompt will use the new version.
Next to the memory file there’s a Manage Edits screen. This is where Cowork stores one-off memories — small facts it picked up mid-conversation. Skim this periodically. You’ll find noise, outdated client names, things you said once as a joke. Prune freely.
Referencing an old conversation
Memory is one way to carry context forward. The other is directly pulling in a past chat.
Inside any Cowork window, you can search your chat history and reference a specific conversation. This is useful when:
- You solved a problem six weeks ago and want to reuse the approach
- A past thread had important reference material you don’t want to re-upload
- You want to move a conversation into a project (we cover projects here)
Search for a topic — for example, “style guide” — and Cowork will list matching chats. Click in, read, and either pull the content forward or move the chat into a project.
What memory is not good for
Memory is “best effort.” It is not the right place for:
- Things that must be precise (quote formats, client codes, file paths)
- Things that must never be forgotten
- Rules that must be applied on every single prompt
For those, use global instructions — covered in the next module. Global instructions are authored by you and are injected verbatim, with no guessing.
Treat memory as a helpful background process, not a trusted system of record.
Common mistakes
- Not looking at the memory file. Most users never open it. Doing so once a month is worth the five minutes.
- Trusting memory to handle firm-specific rules. If it absolutely has to happen every time, put it in global instructions, not memory.
- Letting stale memories rot. Old client names, old roles, old preferences — Cowork will keep pulling them in unless you prune.
Next up
Now that you’ve seen how Cowork remembers you passively, the next thing to learn is how to tell it — explicitly and consistently — what to do every time. That’s global instructions, the short text block Cowork prepends to every prompt you send.