70% of people have walked away from an AI chat mid-conversation because it did not remember something they had already told it. That is not a niche annoyance. It is the finding of Twilio’s 2026 Customer Insights Series, a survey of 7,652 consumers and 660 business leaders across 18 countries, and it lands on a specific and useful gap: 84% of businesses believe their AI recognises a returning customer and picks up where things left off, but 70% of the customers on the other end say they regularly have to start from scratch anyway.

We run on Claude every day, for research, drafting and the daily build behind this blog, so we felt that gap ourselves before we read the statistic. This week we went through Claude’s Projects and Cowork setup properly, closed the obvious holes, and worked out what the equivalent tools in ChatGPT and Gemini do differently. Here is what we found, and how we would set it up if you are starting from nothing.

What actually changed

Claude’s chat and Cowork used to keep separate memories, so something you told it in one did not show up in the other. Anthropic merged the two in mid-September 2026: Projects and Artifacts are now shared across both, and each Project keeps its own isolated memory space and a running summary, kept apart from every other Project and from chats outside one. In practice that means a Project remembers your client, your job or your recurring task, rather than the whole account remembering everything at once and mixing it all together.

That distinction matters more than it sounds. One long-running chat thread for everything is exactly how the 84% versus 70% gap happens: the business assumes the AI has kept up, and it has, just with the wrong client’s details on top.

Set up a Project so it stops forgetting

The fix is not complicated, and none of it needs a developer.

• One Project per job, not one chat for everything. Open a separate Project for each client, each recurring task, or each part of the business you keep re-explaining. Nothing bleeds between them.
• Load your repeat documents once. Your price list, service area and standard terms go into the Project’s files a single time. Every future chat in that Project already knows them, instead of you retyping the same paragraph on a Tuesday and again on a Friday.
• Use the shared memory on purpose. Start a task on your phone in Cowork, pick it up later at your desk in chat, on the same Project, and it should carry straight over. If it does not, the Project needs another pass at its files or instructions.
• Test it before you trust it with a client. Close the chat, open a fresh one in the same Project, and ask it something only an earlier conversation would know. A correct answer means the setup works.

What ChatGPT and Gemini do differently

We are not going to pretend Claude is the only tool that remembers, because the honest answer is more useful than the tidy one, and a fair comparison is what actually helps you pick.

OpenAI rebuilt ChatGPT’s memory this year under the name Dreaming. A background process reviews finished conversations and rewrites its own memory summary afterwards, with no need to say “remember this.” OpenAI’s own benchmark put factual recall task success at 82.8% under the newest version, up from 41.5% on the 2024 system, a real jump if you take it at their word.

Gemini works the other way round. A Gem is a fixed set of instructions plus whatever knowledge files you attach, and it stays exactly as you built it. Tell a Gem something new mid-conversation and it will use it in that chat, but it will not remember it the next time you open the Gem unless you go back and edit the instructions yourself. Google runs a separate account-level memory feature alongside Gems, and its own support forums have carried a steady stream of threads through 2026 about that memory not reliably reaching a Gem’s own conversations.

None of the three is simply better. A tool that rewrites its own memory saves you the admin, but you have less visibility into exactly what changed and when. A tool that only remembers what you explicitly wrote down is slower to keep current, but you always know precisely what it knows. Pick the shape that matches how carefully you want to control what an AI keeps about your business.

The part nobody puts on the tin

Remembering more also means storing more. Whatever a Project, a Gem or a memory summary holds onto is also whatever that tool is keeping on a server somewhere, which is exactly why every one of these systems ships a settings page where you can see, edit or delete individual entries. We would rather you open that page once than find out what it has stored by accident. Anything you would not want kept for months should not go into a chat you expect to persist.

Why this is worth an afternoon, not a project plan

UK small business owners already lose an average of 8 hours a week, 384 hours a year, to admin, and 47% have never tried to automate any of it, according to a March 2026 survey of UK owner-operators. Self-reported AI use among UK businesses with 10 or more staff has climbed from around 12% to around 35% since late 2023, per the Office for National Statistics, so the tools are already in the building for most readers. The gap is not access. It is setup.

A Project that remembers your business does not eliminate that admin pile, but it stops you rebuilding the same context from zero every time you open a new chat, which is where a meaningful share of that 8 hours a week actually goes. Pick the one Project you would benefit from most, load its files this afternoon, and test it with the same question twice, a day apart. If the second answer is better than the first, it is working.