#ChatGPT suddenly forgetting the entire conversation's context

1 messages · Page 1 of 1 (latest)

pearl meteorBOT
#

Reported by @glossy sedge

Bug Report: Non-reasoning models suddenly forgetting the entire conversation's context early in conversations
`Steps to Reproduce`

I don't know exactly how to reproduce it, but I could consistently get this to happen in conversations with multiple file attachments. (Edit: I also noticed that in the conversations I had with this bug, it also happened after 10 messages regardless of whether the files were images or text files, I still don't know if the file attachments have anything to do, but it at least makes this bug easier to notice)

`Expected Result`

The AI to actually keep track of what I just sent it. If it had something to do with the context limit, it should forget the earliest stuff, not the latest.

`Actual Result`

At one point, it suddenly changes its writing style much earlier than usual in the conversation, which normally only happens after reaching the context limit. When I ask it if it remembered what I just showed it, it got everything wrong. This happens regardless of custom instructions or model selected (both GPT-5 and GPT-4o gave the same results). However, if GPT-5 decides to use reasoning, it can remember the earlier context of the conversation, so maybe this only affects non-reasoning models.

`Environment`

Tested in the ChatGPT Windows app and website versions.

#
Additional Information

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

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

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

past ocean
#

I saw something about GPT5 (For plus users) being reduced to 32k tokens. That was super infuriating, because GPT4o had 128K and even 4.1 being capable of 1M, they had it at 32k for plus users.
Originally they had said that GPT5 would have 256K tokens for plus users of context length...

Someone said that GPT5 thinking has 196K vs GPT5-Fast having only 32K.

I may be wrong with this information though... This is just what I have seen and what I have been told here in the OpenAi discord channel

upper mist
#

I'm having this issues as well but for me it's not context window, it's just randomly corrupting and will answer messages from other chats (with reference chats turned OFF), jumping into random things i didn't ask it to that are completely unrelated, or hallucinating that I'd uploaded files in chats I'd never uploaded anything in. I've been using ChatGPT for the last 10 months to work on the same creative writing project. Nothing has changed with the way I use the tool and only the last few days have I noticed this.

When chats get 20-30 messages deep, less then 10k words (WELL below the context limit), it suddenly does this, ignores my message completely and starts talking about random things. Sometimes I can scroll up, edit a message and it fixes it for a bit, but it always gets corrupted further down the chat. 4o and 5 both do this. If I regenerate the message with 4.1, it answers normally. All very weird and annoying as any long-form chat longer than 8k is impossible now.

zinc harbor
#

🚨 For anyone struggling with drift / continuity errors in long threads 🚨

I got so frustrated with canon drift that I built my own memory system workaround.
It’s called the Master Log + Forge Laws framework, and it’s kept my AU stable past 4,000+ inputs without collapsing.

📂 Here’s a plug-and-play bundle:

Fillable Master Log (PDF) → drop your AU Bible/world info in, it becomes your own “memory system.”

README (Quick-Start) → explains how it works in plain language, so you don’t need to wade through the dev spec.

âś… Features:

Locks canon so it can’t drift.

Auto-enforces recaps + snapshots.

Guarantees 4–5 options per choice point.

Keeps meta goals out of the story world.

If this saves even one person from losing their thread, I’ll be glad. Feel free to test it, break it, or adapt it.