If you’ve ever been fifty messages into a project with an AI and watched it contradict something it told you an hour ago, you’re not imagining it. And it isn’t a bug. It’s how these models work.
Here’s the mechanism, why it costs you more than you think, and the workaround.
Every AI chat is stateless
When you send message #50, the model does not remember messages #1 through #49. It has no memory of them at all.
What actually happens: the entire conversation history gets handed back to the model, from scratch, along with your new message. It reads all of it again, then writes a reply. That’s how it appears to “continue” a conversation.
Which means every message you send in a long chat is doing more work than the last one.
The cost curve nobody shows you
Because the whole history is re-read each time, the cost of a single message climbs as the chat grows:
| Chat turn | History re-read each message | Relative cost |
| Turn 1 | ~500 tokens | 1× |
| Turn 10 | ~5,000 tokens | 10× |
| Turn 25 | ~15,000 tokens | 30× |
| Turn 50 | ~40,000 tokens | 80× |
| Turn 100 | ~90,000 tokens | 180× |
| Turn 150 | ~150,000 tokens | 300× |
Turn 50 costs about 80 times more than turn 1 — for exactly the same size answer.
On a subscription, you don’t see this as a bill. You see it as hitting your limit. That’s the real answer to “why did I hit my cap after only 20 messages?” — those twenty messages were inside a chat carrying a hundred turns of context. Long chats eat quota disproportionately.
Context rot: it gets worse, not better
Somewhere around 30 to 50 turns, the AI starts getting worse at your specific conversation. You’ll notice:
- It contradicts a decision it made 20 turns ago
- It forgets file names, character names, constraints you set at turn 3
- It repeats suggestions it already gave
- It gets stuck in a rhetorical groove and can’t break out
- It invents features or decisions that never existed
The reason is attention. Even models with very large context windows don’t read that window evenly — they attend strongly to the start and the end, and weakly to the middle. In a hundred-turn chat, the middle is exactly where your project’s logic lives. So the longer the chat, the more of your important reasoning sits in the model’s blind spot.
The four other things this quietly costs you
Finding decisions. Dozens of chats, most called “New chat”. Which one had the pricing decision? You end up scrolling sidebars.
Model lock-in. You started the project in one AI. Now you want another one’s strengths. But two weeks of history says you’re staying. So you stay — and you’re stuck with whichever model you happened to pick on day one.
Restart cost. Every fresh chat means 5–10 minutes of re-explaining the project. So you avoid starting fresh chats — which is precisely what makes chats too long in the first place.
Interruption cost. Lunch, a meeting, tomorrow morning. You come back and spend five minutes reading your own chat to work out where you were. Three times a day, all week, and it adds up to real hours.
The workaround: stop having long chats
The fix isn’t a better prompt. It’s a different pattern.
Instead of one long chat that slowly degrades, work in short, focused chats — 15 to 25 turns. When a chat has done its job, capture what matters: the decisions you locked, the background that matters, what’s next. Then start a fresh chat and hand it that summary.
A good summary is small — a couple of pages, maybe 500–2,000 tokens. Not 40,000. So your fresh chat starts cheap and sharp, with all your project’s logic sitting right where the model’s attention is strongest.
Done consistently, this cuts token and quota burn on long projects by roughly 60–85%. Not because of any trick — simply because you never let a chat get long enough for the input tax to spiral.
| The honest catch This only works if you actually do it. Most people don’t, because writing a good summary after every chat is work — and because starting a fresh chat feels like losing progress. That’s the real reason everyone defaults to one long chat: not ignorance, friction. |
Making the pattern automatic
This is the problem Jar was built for. You save a checkpoint from your chat into Jar; Jar keeps one canonical Package per project — locked decisions, background, next steps. Starting a fresh chat becomes: open Jar, copy your Package, paste. The AI is oriented in about five seconds.
Because the Package is just text you paste, it works with any AI — ChatGPT, Claude, Gemini, Grok. Which quietly solves the lock-in problem too: use one model for architecture, another for copy, a third to check the numbers, all from the same Package. And if a vendor has an outage, changes pricing, or your account has a problem, your project isn’t trapped inside it.
There’s a free plan that runs the whole loop, and paid plans start at ₹100/month if you need more room. See Jar plans here.
The short version
- AI chats are stateless — the full history is re-read on every message
- So long chats get expensive (or eat your quota) and get dumber, not smarter
- The middle of a long chat is the model’s weakest attention zone — and that’s where your project logic lives
- The fix is short chats plus a good handoff summary
- Do that and you cut token burn 60–85% while every chat stays as sharp as your first one
