OpenAI Community, October 21, 2026: "multiple conversations that previously worked normally have started displaying the following message: Chat memory full."

October 21, 2026 • Reliability Desk • 3 min read

"Chat Memory Full." After Only 9 to 15 Messages.

Since an October 20 app update, users say short ChatGPT conversations hit the memory wall almost at once, the same night ChatGPT, Codex and the API were degraded for about an hour.

A server rack full of tangled network cables

A server rack full of network cables, shown for illustration. Photo: Kim Scarborough, CC BY SA 2.0, via Wikimedia Commons.

Post Facebook Reddit

Wednesday, October 21 brought a wave of ChatGPT users who say their conversations are being cut off almost as soon as they start, a day after OpenAI logged an hour of degraded service across ChatGPT, Codex and its API.

An hour of "degraded performance"

The OpenAI status page shows an incident titled "Degraded performance in ChatGPT, Codex, and APIs" that began at 7:26 PM Pacific on Tuesday, October 20. OpenAI posted that it had "applied the mitigation" at 7:48 PM and marked it resolved at 8:28 PM with "All impacted services have now fully recovered." The incident listed 33 components, among them ChatGPT Conversations, Login, File uploads, Voice mode, Deep Research, Codex Web, the Codex CLI, Chat Completions and Responses.

On Wednesday morning a second incident followed: "Elevated error rates for gpt-5.6-luna API requests," opened at 8:34 AM Pacific and resolved at 9:44 AM, affecting Chat Completions and Responses. A separate notice for "Elevated FedRAMP API error rates for Astra and Sol" was still in monitoring when we read the page.

"Chat memory full" after 9 to 15 messages

On the OpenAI Community forum, a Free plan user on Android wrote that between late October 20 and early October 21, "multiple conversations that previously worked normally have started displaying the following message: Chat memory full." The user tested three conversations and says all three "reached the context limit after only 9 to 15 messages," including a text only chat that stopped at about 15 messages and a continued chat that stopped at about 9.

Another user wrote that it began on October 20 after an app update, with "Completely reasonable limits up to that point then, suddenly, seemingly short conversations" showing the warning, and that "Previously usable, longer conversations have also been retroactively blocked." A reply in the same thread said "many other users appear to be experiencing the same issue." A separate thread asked whether anyone else saw "Chat memory full after 15 mins," and a reply said: "Yep, on the desktop browser. You are not the only one."

A paying user in a third thread wrote that the chat limit is being "reached faster and faster" and that "paid user with chat limit equals to no continuity and disrupting services."

Answers that stop, files that never attach

A ChatGPT Pro user on Windows wrote that GPT 6 "has increasingly failed to complete responses," which "often stop after running for several minutes, sometimes in the middle of a sentence or list," in both the desktop app and Microsoft Edge. Another Pro subscriber wrote that attaching any PDF, PowerPoint or Word file now shows "a constant spinning loading circle" and the file never attaches, "Every time," adding that it "just started today."

On Linux, one user reported the ChatGPT desktop app writing to disk "at high rates (peaking at ~30 MB/s)" while completely idle, and another reported RAM use climbing to 47 to 49 GiB before the app crashed during long agent tasks.

What OpenAI has said

Beyond the status page entries above, we found no OpenAI statement about the memory limit reports. None of the forum threads quoted here had a staff reply when we read them. These are individual user reports; we have not tested any of the affected accounts.

We will update this page if OpenAI responds.

Sources

Keep Digging

← Back to ChatGPT Disaster Homepage