Answers
Shared context across Claude, ChatGPT and Codex
One connection, and every assistant reads the same project. Arroway runs as a server your AI tools connect to — Claude, ChatGPT, Codex and any other tool that speaks the same connection standard — so they all read and write one memory. What you settled in one tool is already there when you open the next, and changing tools does not start you over.
Last updated November 14, 2026
How far a vendor's own memory reaches, in its own words
On 25 August 2026 Anthropic announced that from that date the memory Claude uses in chat is the same one it uses in Claude Cowork, kept as short topic files you read, edit and delete inside those products. The reach that announcement declares is that vendor's own surfaces, tied to your account there; it does not describe reaching any other tool. Arroway sits in the other place by design: the memory lives on a remote server your assistants connect to, and the judgement stays on the server rather than in what you install, so what you settled in one client is read by the next one, whoever makes it. If your work already crosses tools, that gap is what this page is about.
The memory lives on the server, not in what you install
The connection to the server is what every tool gets, and it is enough to read and write the project. For some tools there is also a small plugin, and it does one narrow thing: it reminds the assistant to read at the start and write at the end. Everything that might need to change later — what gets served, in what order, how much — lives on the server, so an improvement reaches everyone at once instead of waiting for each person to update something.
Each person connects as themselves
A team workspace can make the connection available to everyone, but each person still signs in with their own Arroway account. That is what keeps every line in the record attributed to a person, not to a workspace — when a rule shows up in your assistant, you can see who wrote it.
The same material, whatever tool asks for it
The read is put together on the server and delivered to whichever tool asked, in the size that tool can take. A long read arrives in parts, and nothing is silently dropped to make it fit: if a request is too big to honour in full, it is refused whole with the count of what did not fit, rather than answered halfway and called done.
What is not shared, and when this is the wrong tool
The conversation itself does not move: what you said, tried and abandoned inside one tool's chat stays in that chat, and only what was recorded in the project is read by the next tool. That is by design, because a transcript is mostly noise, but it means the discipline of recording is the price of the handover. This is also not a way to make two tools behave identically; each keeps its own style and its own limits, and only the project's decisions are common. If you only ever use one tool and nobody else touches the work, there is no handover to fix, and a plain file in the repository is enough until that changes.
The mistake: treating the second tool's first read as a test of the memory
When the afternoon tool does not seem to know something, the instinct is to paste it and move on. That repairs your afternoon and leaves the record wrong for everyone, including you tomorrow. The better reflex is to ask whether the decision was ever recorded, and if it was, whether the read delivered it — the read says what it left out. Each of those has a different fix, and only one of them is yours to make.
What it looks like in practice
Someone who answered a poll about handing work from one assistant to another put it plainly: the handover between the two tools today is them copying and pasting. Take that as the starting point. A hypothetical founder works on a billing rewrite in the morning with one assistant and in the afternoon with another, and every afternoon begins with the same six pasted paragraphs. The paragraphs are the decisions, the rejected options and the open question. Move those into the project once, and the afternoon assistant reads them itself.
Replacing the copy-and-paste handover
- List what you paste today when you switch tools. For the billing rewrite: the chosen approach, the two options rejected and why, the constraint from the payment provider, the open question about refunds.
- In the morning session, ask the assistant to record those as the work happens, not at the end of the day. It files what you said or approved as decisions, and what it inferred itself as proposals.
- Before you close the morning session, ask what it recorded. Fix one wording that is almost right. This takes about two minutes and is the only review the handover needs.
- In the afternoon, open the other tool and ask it to read the project before it starts. Do not paste anything.
- Compare what the assistant says it knows with your six paragraphs. The approach and the rejected options should be there with their reasons. A gap means a recording gap, so fix the entry rather than pasting around it.
- Work, and close the afternoon the same way, so the morning assistant — or a third tool next week — inherits it. The record, not the conversation, is what travels.
Questions people ask about this
- Which tools can connect today?
- Any tool that speaks the Model Context Protocol, which is the connection standard the main assistants use. Claude has a published connector; ChatGPT and Codex connect through their own settings; other tools and editors connect to the server directly. The installation page shows one route for the tool you actually use, not a list of every possibility.
- Do I have to install a plugin?
- No. The connection alone makes the memory readable and writable. Where your tool supports it, the plugin is recommended: it adds the reminder to read and write at the right moments, so the habit does not depend on you remembering.
- Is context copied between tools?
- Nothing is copied. There is one project on the server, and each assistant reads it directly — so there is no sync that can fall behind, and no two versions to reconcile.
Where this is verifiable
The published Claude connector and the installation page on this site — the two sources a reader can check for themselves. Support per surface is stated only where the client confirms it. The note on native memory repeats only what Anthropic’s announcement of 25 August 2026 declares about itself, and links to it.
https://www.arroway.app/en/answers/shared-context-across-ai-tools