ChatGPT

Connect MemoryRouter to ChatGPT as a custom MCP connector — eligible plans, workspace controls, scopes, and honest limits.

MemoryRouter connects to ChatGPT as a custom MCP connector pointing at the remote MCP server. Your memory follows you between ChatGPT, Claude, Cowork, Claude Code, and Codex.

This is not the Codex integration and not a GPT Action. Codex is the coding CLI and uses hooks for automatic memory. GPT Actions are a different, OpenAPI-based extension mechanism. This page is about ordinary ChatGPT web and app conversations using a connector.

The connector package is version 1.0.0. It has not been submitted to OpenAI for directory review, so it is added as a custom connector rather than installed from a listing.

Setup

Open connector settings in ChatGPT.
Add a custom connector with this URL:
```
https://mcp.memoryrouter.ai/mcp
```
Leave client ID and client secret empty — the server supports dynamic client registration and discovery.
Sign in when prompted and choose which memory vault this connection uses.

You never paste a MemoryRouter API key into ChatGPT. Authentication is OAuth 2.1 with PKCE.

Plans and workspace controls

Custom connector availability is controlled by OpenAI, and on Business, Enterprise, and Edu workspaces an administrator decides whether members may add custom connectors at all. If the option is missing, that is a workspace policy, not a MemoryRouter error — ask your workspace admin to allow it or to add the connector centrally.

Because these controls change, confirm current eligibility in OpenAI's own connector documentation rather than assuming.

What ChatGPT can do

Tool	Requires	Purpose
`search_memories`	`memories:read`	Recall past context
`store_memory`	`memories:write`	Save something durable
`memory_status`	`memories:read`	Show connection, vault reference, counts, scopes
`forget_all_memories`	`memories:delete`	Delete the entire connected vault
`search`	`memories:read`	Compatibility alias

Connections default to read-only. When ChatGPT attempts a save or delete without the scope, the server returns a reauthorization challenge and ChatGPT prompts you to reconnect and approve it — the action does not silently fail, and nothing is written.

Model-directed behavior (important)

ChatGPT decides when to use the connector. There is no hook that forces recall or capture on every turn.

In practice:

Recall is reliable when your message clearly refers to prior context. Being explicit helps: "check my memory for our pricing decision."
Saving generally happens when you ask for it: "remember that we chose vendor B."
Neither is guaranteed on every turn.

If you want deterministic, automatic memory, use Claude Code or Codex, which use lifecycle hooks. Same vault.

Vaults and ChatGPT

One connector connection is bound to one vault, chosen at sign-in. To use a different vault, disconnect and reconnect, then select it.

Personal, project, and organization separation is achieved with separate vaults. Asking ChatGPT to "use the project vault" does not change vaults — the tools take no vault argument. See vaults.

Distinct from ChatGPT's built-in memory

ChatGPT has its own memory. MemoryRouter is a separate, external, OAuth-selected vault, and both can be on at once. The reason to add MemoryRouter is portability: the same memories are available in Claude, Cowork, Claude Code, and Codex, which ChatGPT's built-in memory cannot do.

Privacy and deletion

memory_status returns an opaque vault reference. Your memory key is never revealed to the model or in tool output.
forget_all_memories deletes the whole vault, requires the memories:delete scope, and requires the exact phrase DELETE ALL MEMORIES.
Per-memory deletion is not available through the connector. Remove individual memories in the dashboard.

Limitations

Model-directed: recall and capture are not guaranteed every turn.
Not submitted for OpenAI directory review; added as a custom connector.
Custom connector availability depends on your plan and workspace policy.
One connection is bound to one vault; no in-conversation vault switching.
Whole-vault deletion only.

Troubleshooting

No option to add a custom connector. Plan or workspace policy — contact your admin.

"Insufficient scope" or a repeated reconnect prompt. The connection lacks write or delete authority. Reconnect and approve the scope; retrying without reconnecting cannot succeed.

ChatGPT says it has no memory of something. Ask it to check memory explicitly, then confirm with memory_status that you are on the vault you expect.

ChatGPT

On this page