Memory Tool Breaks When Entries Exceed ~3,500 Words — Incomplete Context, Ghosting, and Errors

[quiet zealot]

Reported by @crimson spruce

Bug Report: Memory Tool Breaks When Entries Exceed ~3,500 Words — Incomplete Context, Ghosting, and Errors

`Steps to Reproduce`

1. On a Plus account with memory enabled, begin adding entries to the Memory tool.

2. Add structured memory entries totaling approximately 3,500–3,700 words (well under the tool’s actual limit of 6,500–7,000).

3. After saving, start a new chat with no system or message context (ChatGPT 4o specifically).

4. Attempt to prompt normally. Observe that behavior begins to break at the ~3,500-word threshold.

`Expected Result`

ChatGPT should incorporate all visible memory entries into a new chat as context, behaving normally until the storage bar is filled. All entries should be readable and affect conversation appropriately.

`Actual Result`

At around 3,600 words, behavior becomes inconsistent:

Memory entries are silently ignored in new chats.

Some conversations start in the middle of prior scenes without prompt.

Some memory entries stop being referenced or integrated entirely, even though they appear saved.

Attempting to save more entries beyond this point leads to message stream error issues or empty responses.

Re-adding entries one at a time causes other entries to “ghost” or become non-functional.

`Environment`

Mobile app, Safari Browser, Google Chrome browser

Jul 29, 2025, 4:47 PM

Additional Information

Please provide relevant details to help resolve the issue, such as:

• ChatGPT Shared Link (if applicable).
• Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

[crimson spruce]

Model affected: GPT-4o

Other models tested: GPT-4.1 and GPT-4.1-mini are unaffected — they behave normally even with memory enabled.

Devices/Browsers tested: Issue persists across all platforms — mobile app, Safari, PC browser, and incognito mode.

Memory size at failure: Issues begin well before the memory warning threshold appears. Breakdown starts at around 3,624 words (roughly 50% capacity), even though total memory capacity is ~6,449 words (93%).

Symptoms:

Model responds with unrelated or mid-scene text in brand-new chats.

Context is dropped or misunderstood.

Memory entries stop being “read” or used in replies.

Attempts to add more memory cause the model to ghost entries or randomly “forget” others.

Eventually, messages return errors like error in message stream or “last few messages are empty.”

Temporary workaround: Turning memory off fixes the issue immediately — model behaves normally again.

[crimson spruce]

I've just tried it on a different browser on my computer and now I'm consistently getting error in message stream notifications. Nothing works now unless I turn the memory off or only have a few entries stored.