Docs

ChatGPT and Claude (Remote MCP)

Connect ChatGPT and Claude on the web to Agent Memory, save the current conversation, and reconnect when permissions change.

ChatGPT and Claude on the web reach your memory through Remote MCP, a connector that signs in with OAuth 2.1. They only see your own data.

Connect

In the connector settings of ChatGPT or Claude, add a custom connector.
Enter the server URL https://agent-memory.jp/mcp.
Sign in and approve access when the authorization screen opens.
Check that the Agent Memory tools appear in the tool list.

Server URL

https://agent-memory.jp/mcp

Tools

Tool	Access	What it does
`memory_search`	read	Search memories by meaning and words. Results include provenance; filter by kind or workspace.
`memory_get`	read	Fetch one memory with its workspace and sources.
`session_replay`	read	Replay the original session: conversation, tool results and recorded thinking.
`health`	read	Check access and count your workspaces, sessions and memories.
`web_session_capture`	write	Save the current ChatGPT or Claude conversation in one call.
`chatgpt_session_capture`	write	The earlier ChatGPT-only version of the above, kept for compatibility.
`memory_add`	write	Add a memory of a given kind.
`memory_feedback`	write	Record whether a memory helped: accepted, rejected or corrected.

Save the current conversation

Ask the assistant to save the conversation, for example “Save this conversation to Agent Memory”. It calls web_session_capture once and returns a session ID that session_replay can open.

What is saved is what the model can see in the current conversation. Sidebar history, other chats, hidden reasoning and messages that have dropped out of the context are not included.

Semantic search

Semantic search over Remote MCP needs a query-embedding key of your own; there is no shared key. Until you add one, search matches words. A query_embedding_not_configured status is expected and means nothing is broken.

When new permissions are added

If Agent Memory adds an OAuth scope, existing connectors don't pick it up on their own. Refreshing the tool list re-reads the tool definitions with the current token; it doesn't authorize again. To update:

Disconnect Agent Memory in the connector settings.
Connect again with https://agent-memory.jp/mcp.
Complete the authorization.
Check that the new tools appear in the list.
Call web_session_capture once and open the returned session with session_replay.

If you still see only read tools after memory:write was added, the connector was not authorized again. Disconnect and reconnect instead of refreshing the tool list. Do this for ChatGPT and Claude separately.