Guide
ChatGPT “Chat memory full”: what it means and what to do
Two different ChatGPT messages talk about memory being full, and most advice mixes them up. Here is how to tell which one you have, how to fix each, and how to stop running into them.
Written October 21, 2026. The “Chat memory full” message is new and OpenAI hasn’t explained it yet; we’ll update this page when it does.
Which message do you have?
- “Memory is full” (or “Memory full” next to a memory update): your saved memories list is at capacity. These are the facts ChatGPT keeps about you across all chats. New things won’t be saved until you make room. Your conversation itself works fine.
- “Chat memory full… start a new chat”: one conversation can’t continue. People started reporting it in October 2026, sometimes after only a short chat, on OpenAI’s community forum. Some versions offer to start a new chat with an automatic summary, or to upgrade.
Fixing “Memory is full”
- Back it up. Ask ChatGPT to list everything it remembers about you, and save the answer in a document.
- Open Settings › Personalization › Manage memories.
- Delete in this order: finished projects, one-off facts (“is planning a trip to Lisbon”), duplicates and near-duplicates, then preferences you no longer hold.
- Merge the rest. Ask ChatGPT to rewrite several related memories as one, then delete the originals.
Deleting chats doesn’t free memory: saved memories are stored apart from conversations. OpenAI’s memory FAQ covers the settings.
Fixing “Chat memory full”
You can’t clear space in that conversation. The fix is to move to a new one without losing the thread:
- Before you hit the wall (or right after, if ChatGPT still answers), ask for a handover. The prompt is below.
- Read it. Fix anything wrong; it’s now the only record of the conversation’s conclusions.
- Start a new chat and paste it first.
- If you use Projects, keep the handover in the Project’s files so every new chat starts with it.
The automatic summary ChatGPT offers is better than nothing, but it’s written for continuity, not for decisions. A handover you check is more reliable.
A handover that works
Write a handover for continuing this conversation in a new chat,
under 300 words:
- Goal: what we're trying to get done
- Decided: what's settled, with the reason
- Constraints: limits, requirements, preferences that matter
- Ruled out: options we rejected and why
- Next: the very next step
Only include what was actually said here. Mark anything uncertain.“Ruled out” is the line people skip, and it’s why a new chat suggests the thing you rejected an hour ago.
What’s known
As of October 21, 2026, OpenAI hadn’t published an explanation of the per-chat message, and forum threads about it had no staff replies. It may be a new limit, a test, or a bug. Until OpenAI says, assume long or file-heavy chats can hit it, and keep handovers as you go on anything important.
Stop storing work in chat memory
Both messages come from asking chat memory to hold your work: projects, clients, decisions, what happened this week. It’s a poor place for that. Saved memory is a short list built for preferences, and a conversation is a scratchpad.
Work context holds up better outside the chat:
- A Project with a brief you update weekly. Simple and reliable if you keep it up.
- A memory connector that records the work for you. Cachejay captures screen text, calls and coding-agent sessions on your Mac and files them by project. ChatGPT can read it on Business, Enterprise and Edu plans, and read-only on Pro. Plus and Free can’t add custom connectors, so on those plans use it from Claude or another MCP app.
More in stop pasting context into ChatGPT and why AI memory goes stale. If you’re thinking of leaving, see switch from ChatGPT to Claude.