ChatGPT Memory: assistant inconsistently claims direct memory updates vs. correctly stating persistence is asynchronous

Summary

ChatGPT's conversational handling of Memory capabilities is internally inconsistent. In some conversations the assistant behaves as though it can directly save/update a user memory and may state that the memory was updated; in other conversations it correctly states that persistent Memory updates are asynchronous and that the assistant cannot directly force-write the memory store from the current turn.

This creates a misleading product contract for the user: the same assistant alternates between implying a direct write capability and denying that capability.

Observed behavior

In a current ChatGPT web conversation, the user explicitly asked the assistant to write a communication preference into memory.

The assistant responded with the intended preference, but then said:

"A memóriarendszerbe ezt innen nem tudom kézzel, azonnal 'beírni'; a tartós memória frissítése aszinkron történik."

The user then pointed out that in other conversations the assistant sometimes spontaneously reports the equivalent of "memory updated" as though it had directly performed the write.

This is not merely a wording issue. It affects user expectations about:

• whether a requested memory change has actually been committed;
• whether the assistant has a direct memory-write capability;
• whether the user needs to repeat important preferences later;
• whether a confirmation from the assistant is authoritative or just conversational language.

Reproduction

1. In ChatGPT, ask the assistant to remember a durable preference or fact.
2. Repeat this across separate conversations / sessions.
3. Observe inconsistent capability claims:
   • in some cases the assistant may present the change as completed ("memory updated", "I'll remember that", etc.);
   • in other cases it says it cannot directly write persistent memory and that memory creation/update is asynchronous.
4. Ask the assistant what capability it actually has. The answer may contradict prior behavior.

Expected behavior

The assistant should expose one consistent user-facing contract for Memory.

If the assistant cannot synchronously commit a persistent memory, it should never claim that a memory was definitively written or updated.

If there is a tool/event that performs an actual write, the assistant should only confirm success after receiving a successful result from that mechanism.

Suggested invariant:

No successful memory-write acknowledgement without a successful memory-write event/tool result.

For asynchronous memory systems, safer wording would distinguish:

• "I will use this preference in this conversation / future responses when available"
  from
• "This has been persisted to Memory."

Why this matters

This is an instruction-to-state / state-reporting reliability problem similar to other cases where the assistant verbally claims a state transition that it cannot verify.

For Memory specifically, false-positive confirmation is especially harmful because the user may stop repeating an important preference under the assumption that it has been stored.

It also makes the system appear unreliable even when the underlying memory backend may be functioning correctly, because the assistant's description of its own capabilities is inconsistent.

Environment

• ChatGPT web
• GPT-5.6 Sol
• Observed 2026-09-19

The report names no repository files, tests, or entry points and describes behavior in the ChatGPT web product rather than the Codex codebase. First determine whether Memory behavior is implemented here; done would require a repository location and a reproducible test or verification path for consistent write acknowledgements.