Have a long coding chat below, then ask about the very first thing you said.
- The app's hidden instructions~533 tok
- package.json + config~1,333 tok
Everything still fits. Keep going and watch what happens when it doesn't.
Token counts are estimates (~3 characters per token); real tokenizers differ by model.
It's gone. Not deleted from your screen, but no longer on the model's desk. The [context window](/what-is-a-context-window-2026) moved on.
The model doesn't remember your chat. Each time, the app sends it the conversation again. Anthropic's docs describe each turn's input as "all previous conversation history plus the current user message".
When the history gets too long, something has to go. Anthropic notes that chat interfaces "can also manage the context window on a rolling 'first in, first out' basis." Oldest first.
Three ways a long chat loses detail
Trimming
- What happens
- The oldest messages drop off the front
Summarising (compaction)
- What happens
- Old messages are replaced by a shorter summary; detail is lost
Context rot
- What happens
- Everything is still there, but accuracy drops as it grows
Summarising has a name. Anthropic calls server-side compaction "the primary strategy for context management" for long-running conversations.
So when a vibe-coding session suddenly ignores a decision from an hour ago, it usually isn't stubbornness. The decision fell out, or got summarised away.
Try it with the rules off. In the sandbox, switch trimming off and keep chatting until the window overflows. Then switch summarising on and see which details survive the summary.
Check yourself
0/4 got itSaved on this device only. No account, no streaks.
