Skip to content
OceanDB

ChatGPT MCP memory server

Persistent external memory for ChatGPT with MCP

Connect eligible ChatGPT plans to an external, writable MCP memory that can also be used by Claude, Cursor, and Codex.

Client
General AI assistant · web
Availability
Plan-dependent custom MCP support
Markdown

ChatGPT custom app setup

Connect OceanDB to ChatGPT 

https://app.oceandb.ai/api/mcp
  1. 1. Confirm plan and workspace access. Full MCP write actions currently depend on ChatGPT plan, workspace policy, and developer-mode availability. A workspace admin may need to enable custom apps first.
  2. 2. Create a custom app. In ChatGPT settings, open Apps, enable developer mode where required, choose Create, and enter the OceanDB remote MCP endpoint.
  3. 3. Scan tools and authorize. Let ChatGPT discover the server tools, complete OceanDB OAuth, and review the requested capabilities before enabling the app.
  4. 4. Enable OceanDB in a chat. Select the app for a normal chat, call onboarding_status, and test both remember and recall. Availability in agent mode and deep research differs from standard chats.
Check the current ChatGPT MCP documentation

External memory and ChatGPT memory solve different problems

ChatGPT’s built-in memory can personalize the ChatGPT experience. OceanDB is useful when the memory must remain visible and portable outside that product boundary.

The external store has a documented schema, personal and shared workspaces, a raw log, a maintained wiki, typed relationships, and export to Markdown or SQL. The same facts can be recalled by another MCP client after the user authorizes it.

This makes OceanDB complementary to native memory rather than a claim that ChatGPT has no continuity of its own.

Check capability before relying on write access

ChatGPT’s custom MCP support is evolving. As of this page’s verification date, OpenAI documents full MCP—including modify and write actions—as a beta for Business and Enterprise or Edu plans. Pro access is more limited, and workspace administrators control custom app availability in managed plans.

OceanDB needs write tools for the complete memory loop. A read-only connection can retrieve context but cannot capture new entries. Confirm that remember, recall, execute_sql, and onboarding_status appear during the tool scan, and test a write before treating the integration as complete.

OpenAI also distinguishes normal chats from other modes. Its current documentation says agent mode does not use custom apps, while deep research can use them only for read or fetch actions. Use a normal eligible chat when testing capture.

What to keep in external memory

Good cross-agent memories include:

  • a preference that should also shape Claude or Codex;
  • a decision made during research that Cursor will implement;
  • a project constraint that spans several conversations;
  • a correction that should survive a model or client switch;
  • an open question another agent can continue.

Avoid bulk-copying every ChatGPT transcript. Durable memory works best as a selected record of facts and decisions, with provenance and workspace scope.

Test write and recall separately

In a normal chat with the OceanDB app enabled, ask:

Remember that technical explanations should lead with the decision and then show the evidence.

Start a new chat with OceanDB enabled and ask for the explanation preference. Confirm that ChatGPT invokes the memory tool rather than merely guessing from native personalization.

Then connect Claude Code, Codex, or Cursor and ask the same question. Cross-client retrieval is the proof that the fact belongs to the external store.

Keep governance outside the chat

ChatGPT can capture and recall through MCP, but OceanDB keeps human review in its own dashboard. The agent-facing database role cannot delete, truncate, or change the schema. People can inspect entries, edit maintained pages, archive content, review contradictions, and export the store.

That division preserves a useful boundary: the assistant can work with memory, while the memory remains governable outside the assistant.

Common questions

Is OceanDB the same as ChatGPT's built-in memory?
No. Built-in memory belongs to the ChatGPT product experience. OceanDB is an external MCP store with its own schema, dashboard, workspace permissions, maintenance process, and Markdown or SQL export.
Which ChatGPT plans can write to a custom MCP app?
OpenAI currently describes full MCP write support as a beta for Business and Enterprise or Edu workspaces, with plan-specific developer-mode behavior. Check the current ChatGPT app documentation before setup because availability is changing.
Can ChatGPT agent mode use OceanDB?
OpenAI currently states that agent mode does not use custom apps. Deep research can use custom apps only for read or fetch actions, while normal eligible chats can use approved write actions.
Can ChatGPT share this memory with Claude or Codex?
Yes. When each client authorizes as the same OceanDB identity, it reaches the same personal and shared workspaces through the remote MCP endpoint.

One memory, ready in every agent.

Connect ChatGPT, capture the first durable fact, and keep the store under your own identity.

Start remembering