The most simple explanation I’ve heard for how an AI remembers things goes like this: Imagine a long, narrow table, and then imagine that you begin placing dominos on one end, slowly sliding them toward the opposite end as you add more.
So, you keep chatting and chatting, placing more dominos onto the table while the very first dominos you added slide closer to the other end. If the conversation goes long enough, you’ll start hearing the clatter of the oldest dominos falling to the floor, as they’re nudged off the table by the newer dominos.
That’s a very rough description of how context windows work in AI. Today’s LLM models have limited context windows that can fill all the way up — and when that happens, something has to go, namely the very beginning of your chat.
Now, the ChatGPT, Claude, and Gemini apps all have methods for dealing with context windows that get too full, automatically “compacting” a conversation that’s gotten too long. Claude, for example, will occasionally pause a chat while it crams the context down to a more manageable size.
But even with the best AI tools, you can’t ever tell precisely when they’ll decide to compact a lengthy chat, and while some AI apps will let you trigger a “compress this chat” session (Claude Code and Codex come to mind), mainstream AI applications like ChatGPT typically don’t.
That said, there is a way to manually compact a long AI conversation and then hand the distilled nugget to a fresh chat, essentially rebooting the thread (while saving precious AI tokens in the process).
Now, you could simply ask ChatGPT something like this: “Summarize our conversation so far into a tight 100-word paragraph,” but that risks losing important nuances, including critical forks in the road or even the original problem you’re trying to solve.
Instead, try this prompt (which distills similar “compact this chat” prompts floating around the web):
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.
This prompt should yield a brief overview of a long conversation that nonetheless carries over the most important nuggets of information, perfect for teeing up a fresh new chat.
I tried this prompt for a lengthy discussion I was having with ChatGPT about porting over a classic Apple II role-playing game to a Raspberry Pi 5 (yes, it’s another one of my nerdy projects).
Here’s the response I got:
We’re exploring how to get the original Apple II Wizardry: Proving Grounds of the Mad Overlord running legally on a Raspberry Pi 5, likely through Apple II emulation, not by rebuilding the game from scratch.
Key decisions so far: the Pi 5 is easily powerful enough; the main issue is not hardware but legal access to the original Apple II disk/software. A faithful clone or remake would be possible in Godot/Pygame, but the immediate goal is the authentic original Apple II version.
Important caveat: don’t assume “abandonware” downloads are legal. Wizardry is still a live IP, and random .dsk/.woz disk images online are probably unauthorized. The cleanest legal options are either buying the modern licensed remake for gameplay, or buying an original physical Apple II copy and imaging the disks yourself. The latter is the most authentic but may require specialized hardware such as Applesauce or KryoFlux, and old disks may be degraded or copy-protected.
Very next step: decide whether the goal is authentic Apple II emulation on the Pi or simply legal gameplay. If authentic emulation, research current availability/prices for original Apple II
Wizardry disks and what hardware/software is needed to image them into a usable disk image.
Then I handed off the conversation to a new chat thread like so:
Let’s continue this chat, here’s where we left off: [Paste in generated summary]
Give it a try the next time your ChatGPT conversation gets too long, and you’ll help prevent any key context “dominos” from slipping off the table.