[BUG] ChatGPT connector repeatedly prompts for permission and loses capability between discovery and execution

[TeaShaman-cyber]

Summary

Basic Memory Cloud is intermittently unreliable when used through ChatGPT. The connector can be visible, its tool schemas can be exposed, and a write can be prepared — but the capability may then be disabled or disappear before execution. Repeated permission prompts also interrupt already-authorized workflows.

This is not a one-off call failure. It has recurred across multiple ChatGPT conversations since mid-July 2026.

Environment

• Client: ChatGPT
• Integration: Basic Memory Cloud via MCP connector
• Account type: Basic Memory Cloud
• Approximate incident window: 2026-07-14 through 2026-07-31
• Latest fresh reproduction: 2026-07-31 around 13:17 Europe/Berlin

Observed behavior

Typical failing sequences:

tool namespace visible
→ tool schema visible
→ operation prepared
→ permission/runtime gate
→ capability disabled or disappears

and:

connector works in one chat
→ new chat asks for permission again
→ authorization spinner may not complete
→ connector later disappears

Latest reproduction:

Basic Memory tools discovered
→ write_note schema exposed
→ canonical write prepared
→ runtime response: "The Basic_Memory_Cloud tool has been disabled. Do not send any more messages to Basic_Memory_Cloud."

This reproduced the exact discover → prepare write → capability disabled before execution failure immediately after a detailed report had been emailed to support.

Expected behavior

Once Basic Memory Cloud is authorized:

1. Authorization should persist across ordinary new chats unless the token expires or permissions materially change.
2. Read-only operations should not repeatedly trigger interactive permission prompts.
3. Tool discovery and subsequent execution within the same workflow should use a stable capability session.
4. Failures should distinguish states such as:
   • authorization required
   • token expired
   • permission denied
   • provider unavailable
   • client/runtime disabled
5. Async writes should expose a stable operation ID and observable states such as:
   • accepted
   • processing
   • indexed
   • searchable
   • failed

Actual impact

The issue breaks the critical persistence workflow:

retrieve context
→ decide
→ write durable result
→ read back
→ verify persistence

When the connector disappears mid-workflow, the agent cannot tell whether the write was never sent, accepted, completed, failed, or became inaccessible.

As a result, Basic Memory currently has to be treated as DEGRADED for interactive hot-path use, despite being valuable as a Markdown-based canonical memory layer when calls succeed and readback verifies persistence.

Timeline / control observations

• 2026-07-20: repeated permission prompts paused generation despite broad permissions already being granted.
• 2026-07-21: Basic Memory namespace and schemas were visible in one runtime, but actual dispatch returned a disabled state while other connectors remained executable.
• 2026-07-24: several important Basic Memory writes and exact readbacks succeeded, showing the account/workspace are not permanently broken.
• 2026-07-27: HTTPS MCP route appeared operational with OAuth 2.1 + PKCE and existing authorization state.
• 2026-07-31: all connectors temporarily disappeared from the ChatGPT runtime; later, Basic Memory was rediscovered, but the capability was disabled exactly when write_note was invoked.

Root cause status

UNKNOWN. The failure may be in Basic Memory, ChatGPT's connector permission/runtime layer, or the interaction between them.

Questions

1. Does Basic Memory intentionally require separate permission for fetch, build_context, graph traversal, large note bodies, or full-document retrieval?
2. Are MCP tools annotated with read-only/destructive/idempotent/open-world hints in a way ChatGPT can consistently interpret?
3. Are there known issues with OAuth consent persistence, PKCE refresh, dynamic client registration, or connector disappearance after discovery?
4. Can Basic Memory expose diagnostic state such as token expiry, connector session ID, last successful request, last failed request, and provider correlation ID?
5. Is there a recommended low-friction ChatGPT pattern such as:

search_notes
→ exact fetch only when needed
→ write_note
→ exact readback

Privacy

No API keys, OAuth tokens, private note contents, billing details, or workspace secrets are included here. Sanitized screenshots, exact tool names, timestamps, and private workspace details can be provided through support email if needed.

A detailed report has also been sent to support@basicmemory.com.

Thank you for building Basic Memory. Its local-first, user-owned Markdown model is unusually aligned with long-term human–AI collaboration; this issue is about making that connector path reliable enough to trust.