There’s a simple way to think about how an AI remembers things in a long conversation. Picture a narrow table where you place dominos on one end, pushing them toward the other side as you add more. Keep chatting, and the oldest dominos eventually fall off the edge. That’s roughly what happens when a large language model’s context window fills up — something has to go, and it’s usually the beginning of the chat.
ChatGPT starts forgetting your early messages as the context window fills up
ChatGPT, Claude, and Gemini all have ways of handling context windows that get too full. They’ll automatically “compact” a long conversation, squeezing it down to a more manageable size.
Claude pauses a chat while it crams the context together.
But you can never predict exactly when that compaction will happen. Some AI tools let you trigger a compression manually — Claude Code and Codex come to mind — but mainstream apps like the service typically don’t offer that button.
Related: Old Androids Find New Life as Routers
There is a workaround. You can manually compact a long conversation and hand the distilled version to a fresh chat, effectively rebooting the thread while saving tokens.
A prompt that actually works for compacting long threads
You could just ask the AI to summarize the conversation into a tight 100-word paragraph. But that risks losing important details — forks in the road, the original problem, key decisions. A better approach uses a more structured prompt.
Here is the prompt, adapted from similar online versions:
“Create a handoff summary I can paste into a new chat. Include: what we’re trying to accomplish, key decisions we’ve made, anything important you’d get wrong by guessing, and the very next step. Keep it short enough to read in under a minute.”
Related: How Reliable Are Surge Protectors
The result is a brief overview that carries over the most important nuggets, ready to paste into a clean thread.
How it played out with a real project
I tried this on a lengthy discussion about porting an old Apple II role-playing game to a Raspberry Pi 5. The generated handoff summary captured the goal, the key decisions, a critical caveat about legal access to the original disks, and the next step.
It noted that the Pi 5 is easily powerful enough, that the main issue is legal access to the original Apple II software, and that “abandonware” downloads are probably unauthorized.
It flagged clean legal options like a licensed remake or disk imaging.
Related: What are Popular Custom Business Application Development Tools?
Then I handed off the conversation to a new chat thread with a simple line: “Let’s continue this chat, here’s where we left off:” followed by the pasted summary.
The technique isn’t perfect — it’s a bit of a hack, but it works.
It prevents them from slipping off the table entirely.
Give it a try the next time your AI conversation gets too long. You’ll keep the key context intact without relying on the AI’s automatic compaction schedule.