Claude Memory Now Spans Chat and Cowork

On Tuesday 25 August, Anthropic connected two things that had been living in separate rooms. The memory Claude builds inside an ordinary chat window now carries into Cowork, the mode that runs longer cloud-based tasks, and anything Cowork picks up while it works flows back into chat. One shared memory instead of two that kept forgetting each other.

Anyone who has opened a Cowork session and burned the first five minutes re-explaining what their company does, that their co-founder loathes bullet points, and that the weekly report always goes to the same three people will understand the appeal. That re-explaining tax is a genuine drag on getting anything useful done, and Anthropic has decided to bin it.

Table of Contents

What actually changed

Before this, Claude’s chat memory and Cowork ran on separate context. You could set out your preferences in one place and watch it draw a total blank in the other. Now there is a single pool of editable, topic-based memories that both sides read from and write to.

A few specifics worth knowing:

It is on by default for consumers on the Free, Pro and Max plans, across web, desktop and mobile.
Claude adds topics to memory while you are still talking, rather than waiting for the conversation to end.
There is a dedicated Settings tab and a memory browser where you can view, edit and delete what it has stored, plus tell it what to remember directly.
Health, politics and gender identity stay out of memory unless you explicitly opt in.

That last point is the sensible bit. Sensitive-topic toggles default to off, so Claude is not building a dossier on your medical worries after a single late-night vent.

Why persistent memory is the whole game for agents

Memory sounds like a convenience feature. It is closer to the load-bearing wall for everything Anthropic wants Cowork to become. An assistant that already knows your project status, your recurring priorities and how a specific colleague likes updates written can move from ‘tell me something’ to ‘go and do the thing’ without a fresh briefing every time. Strip the context out and every agent task starts from zero, which is exactly why long-running agents have felt so clumsy.

Google, OpenAI and the rest are all chasing the same target: an assistant that accumulates context across months rather than minutes. Anthropic’s version is notable for crossing the chat and Cowork boundary in both directions, which is the join most people actually hit.

The bit worth worrying about

On by default is a choice, and it cuts both ways. More stored context means more chances for a stale or plain wrong fact to lodge in memory and steer later answers without you noticing. If Claude decides in July that you have left a job you actually still hold, that error can follow you around until you go digging in the memory browser to correct it.

Reliability matters more now too. The more you lean on one assistant to hold your working context, the more it stings when the service falls over, and shared memory raises those stakes. ‘On by default across web, desktop and mobile’ also means the honest homework is boring but real: go into that Settings tab and check what it thinks it knows about you.

Used well, it is a proper upgrade. The friction it removes is the friction that made Cowork feel like starting a new job every morning. The catch is simply that a memory you never audit is a memory you have handed the keys to.

Did you know: information theory, the maths that underpins how any of these models compress and store what they know, dates to a single 1948 paper by Claude Shannon, who also gave us the word ‘bit’.

Sources

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →