LLM 0.32rc1 Lands Content-Addressable Logs That Deduplicate Repeated Conversation Prefixes by Message-Part Hash
Simon Willison·medium signal
Simon Willison shipped LLM 0.32rc1 on 2026-07-30, completing schema work begun in 0.32a0 with a redesign that better captures prompt and response detail. The mechanism that matters: OpenAI-style clients resend the entire growing message array on every turn, and the new schema 'de-duplicates these using hashes of the individual message parts' rather than storing each near-identical array in full. For anyone logging agent traffic to SQLite, this is the difference between linear and quadratic storage growth on long sessions.