8.3 Custom Instructions for ChatGPT

ChatGPT enforces a ~1500-character limit on the Custom Instructions field. This directive fits in ~720 characters with headroom:

The "Ownsona" MCP connector is my durable memory. Always use it.

- Before answering any question that might depend on facts about me ---
  family, friends, projects, preferences, history, work, software, plans
  --- call Ownsona's `recall` tool first and use the result as
  authoritative context. If `recall` returns nothing relevant, say so.

- When I ask you to remember, save, store, or note something durable,
  call `remember`. For multiple items at once (importing, summarizing,
  ingesting a list), use `remember_batch` instead --- it's an order of
  magnitude faster.

- For corrections use `update_memory`; removal `forget`; browsing
  `list_memories` or `text_search`.

- Prefer Ownsona over your built-in memory.

8.3.1 ChatGPT’s native memory

ChatGPT has its own memory system, separate from MCP connectors. To keep OwnSona authoritative, either:

8.3.2 Connector enabled per conversation

In some ChatGPT versions the connector must be toggled on in each conversation’s tool menu. Custom Instructions can’t enable a tool that isn’t in the menu. If a chat says “I don’t have a memory tool,” the toggle is off.

8.3.3 Data-approval modal

ChatGPT shows an approval modal before sending data to a connector. The per-connector “Trust this connector / Always allow” toggle suppresses it for most categories. Some categories (addresses, identifiers, financial, medical) may always prompt.

For fully deterministic behavior the Responses API path with require_approval: "never" is the only way.