Conversation Memory Compactor
Every multi-turn chat resends the entire history on every call — a 40-turn conversation pays for turn 1 forty times. The cost grows quadratically. This engine quantifies the resend tax and prescribes the pruning policy that kills 40-70% of it.
Simulate your conversation shape
Your live history-bloat scan (last 30 days)
The quadratic tax, explained
A T-turn conversation with t tokens per turn and an S-token system prompt costs T·S + t·T(T+1)/2 input tokens if you resend full history — that's quadratic in conversation length. With keep-last-N + a 150-token rolling summary the cost becomes linear. On a 40-turn support conversation at 200 tokens/turn, that's the difference between ~196K and ~78K input tokens — a 60% cut, on every long conversation, forever.
First-to-market: as of June 2026 no FinOps or LLM-observability tool quantifies the history resend tax.
Next: put your responses on a diet