Move the project, not the chat. The work record travels.

What "losing the work" actually means

A project is not its chat. The chat is two participants talking; the project is what the talking produced. When you open a new AI and it knows nothing, what is missing is specific: what was done and the results it produced, what was decided and what was deliberately not done, and the state of play, which step is current and what comes next. Exporting the transcript hands the new AI a retelling of all three, in prose, to take on faith. A work record hands it the steps themselves, each one checkable. The difference matters the first time the new AI builds on a result the old one invented.

What carries across

• The work itself, as receipts. Every call made through the connection left a verifiable receipt when it ran: the tool, the time, the saved result. The receipts are the project's steps, readable from any connected AI or a browser.
• Memory on the connection. Facts saved with memory_remember in one AI are recalled with memory_recall in another: the client's constraints, the chosen approach, the numbers that must not change. Memory belongs to the connection, so it does not care which AI is asking.
• The handover. The first AI packages the project (goal, step receipts, the not-checked list, next steps) behind one link. The second AI opens the same link and continues from the record, and its new steps append to the same project instead of starting a parallel one.
• One shared view. The Today view lines up everything the AIs did in order, so after the move there is still one place where the whole project reads as a single line of work, whichever AI did which part. Which AI did what is that record read per step.
• The decisions. Where work was accepted or flagged, the acceptance state travels with the record, so the receiving side inherits what was already approved, not just what was done.

What does not carry, stated plainly

• The platform's native chat history. The conversation as ChatGPT stores it does not import into Claude, and no outside tool can move it. If you need the conversation, export it from the platform that owns it.
• Another app's internal state. Its own memory files, settings, and workspace state belong to that app and stay there. What moves is what was saved to the connection.
• Work that never ran through the connection. It was never recorded, so it cannot be carried, verified, or recalled. It can be handed over manually as files and notes, and the record starts there.
• Correctness. A verifiable receipt shows the saved record of a step has not changed since it was saved. It does not prove the step was right, on either side of the move.

The honest summary: this moves the work, and the work is the expensive part to lose. Anyone promising to move the chat itself is promising something only the platforms can do.

The move, step by step

1. In the first AI, package the project. With Zambo connected, save the facts that matter with memory_remember, then create the handover: the goal, the step receipts, and the next steps for whoever continues.
2. Open the second AI and add the same connection. One endpoint, https://zambo.dev/mcp, the same one the first AI used. No export file, no conversion.
3. Open the handover and recall the memory. The second AI reads the goal, the steps, and the saved facts from the record itself.
4. Verify before building. Open the step receipts from the second AI and check them. The record reads the same from every client, which is the point: trust is checked, not transferred.
5. Continue. New steps join the same record. The Today view now shows both AIs' work in one line, and the next switch, in either direction, is a non-event.

The concept underneath, handing over the record rather than a retelling, is AI agent context handoff; the agent-to-agent form is handover between AIs, receipts attached; and the session-level version is continuity across sessions. This page is the same idea aimed at the human mid-project, moving between the AIs they already pay for.