Connect Memory V5 to your AI client.
Initial setup takes a few minutes. After that, changing chats is a four-step flow: save the current task, open a new chat, restore the task state, and confirm it before continuing.
Choose how you want to use it.
Standard: MCP only
No browser helper is required. Before changing chats, ask Memory V5 to save the task and prepare it for continuation, then send the generated package into the new chat.
More convenient: MCP + Power
Power is for ChatGPT Web. It prepares the continuation package and copies it. Power does not open a new Chat or send any message for you; you still open the new chat, paste, and send it yourself.
Manual ZIP install for now. Chrome Web Store later.
With MCP only, the flow is four steps.
Connect Memory V5
Add the Memory V5 remote MCP service to your AI client.
Prepare continuation in the old chat
Before changing chats, ask Memory V5 to save the task and prepare it for continuation.
Open a new chat
Send the generated continuation package into the new chat.
Review the restored state and confirm
Check the task goal, current focus, completed work, excluded approaches, unresolved issues, and next step before confirming the restore.
ChatGPT Web
The full connect → save → new-chat restore flow has been tested on ChatGPT Web.
Enable developer mode
Open ChatGPT settings and enable developer mode. Menu labels can vary by account or workspace.
Add Memory V5
{{RESOURCE_URL}}In the custom MCP setup, use this HTTPS server address:
Complete sign-in
If an authorization page opens, follow the sign-in flow. Then return to a real long-running task.
If your ChatGPT account does not show a custom MCP entry, that capability may not be available to your account or workspace yet.
Codex CLI / IDE
Codex supports remote MCP. Memory V5 has not yet completed its full end-to-end continuation test on Codex.
Run:
codex mcp add memory-v5 --url https://memory-v5-beta.8084867.workers.dev/mcp
Then check that it was added:
codex mcp list
If sign-in is requested on first use, complete the authorization flow, then test with a real long-running task.
Gemini CLI
Gemini CLI supports remote MCP and OAuth. Memory V5 has not yet completed its full end-to-end continuation test there.
Run:
gemini mcp add --transport http memory-v5 https://memory-v5-beta.8084867.workers.dev/mcp
Check the connection:
gemini mcp list
If OAuth authorization is required, run:
/mcp auth memory-v5
Complete sign-in in the browser, then return to Gemini CLI.
Other remote-MCP clients
A client needs remote MCP support and a compatible OAuth flow.
Prefer Streamable HTTP / HTTP MCP. Server address:
https://memory-v5-beta.8084867.workers.dev/mcp
If the client cannot complete the required OAuth flow for remote MCP, Memory V5 may not work there yet.
Do not send passwords, cookies, OAuth access tokens, or other secrets.
Install the Power browser helper
Download and extract
Download Power 0.7.0 ZIP and extract it to a fixed folder.
Open extensions
Enter chrome://extensions/ in Chrome.
Enable developer mode
Turn on Developer mode and choose “Load unpacked”.
Select the folder
Choose the extracted memory-v5-power-0.7.0 folder, then open or refresh ChatGPT.
The manual version does not auto-update. Download a new ZIP when a new version is published.
What does Power save you from doing manually?
Prepare continuation with one click
When you are ready to change chats, click Power and it prepares continuation content from the current task.
Copy automatically
When the continuation content is ready, it is copied to your clipboard.
Limited local emergency continuation
During a genuine cloud, network, or sign-in failure, Power can use limited local information to prepare an emergency continuation.
Power does not open a new Chat or send any message for you. You remain in control of the final send.
Test with a real long-running task.
Choose a task that has already progressed
Development, research, writing, product, or operations work is fine. It should already contain real decisions, completed work, and unresolved issues.
Prepare continuation in the old chat
Save this task and prepare it for continuation.
Memory V5 generates the information needed to restore the task in a new chat.
Open a completely new chat
Send the continuation package and review the restored goal, current focus, completed work, excluded approaches, unresolved issues, and next step.
Confirm only when it is correct
Confirm. Resume normally.
If anything is missing or wrong, do not confirm. Point out the problem first.
What counts as a successful test?
In a completely new chat, you can continue from the previous unfinished point without re-explaining the whole project.
If something breaks, include these details.
- Platform and version
- Whether the problem happened during connection, sign-in, save, restore, or confirmation
- Error text or a screenshot
- Whether it could not connect at all, or connected but restored the wrong state
Do not send: passwords, API keys, cookies, OAuth access tokens, or continuation tokens.