As a conversation grows, two things happen at the same time:
The solution is the compaction summary: at some point between turn 10 and 15, pause the conversation, write ONE message that captures the full current state, and continue from there (ideally in a new conversation, or as explicit reference).
A useful summary isn't "a paragraph about what we talked about". It has structure:
If any of the 5 is missing, the summary loses value. Without locked decisions, the model re-litigates them. Without open questions, you lose track of what's missing. Without next step, the summary is just a postcard.
| Signal | Action |
|---|---|
| The model repeats things you already said | Time to summarize. |
| You start forgetting turn 3 | Time to summarize. |
| The conversation passes 12-15 turns | Summarize preventively. |
| Next 3 turns will be a different topic | Compact what's current before switching. |
Advanced trick: ask the model to BUILD the structured summary in 5 slots itself. "Summarize this conversation into these 5 buckets: locked decisions, open questions, constraints, last action, next step." You check it's right, save it, use it as the opening of the next turn (or the next conversation).
On the right, a 14-turn conversation in compressed form. Drag each fact into the right slot.