9.6 Conversation Summaries and User Memory

Trimming keeps a conversation inside its token budget, but it throws information away. The previous lesson stored Amy's travel-planning session abc123 in Redis: a hash sess:{abc123} for metadata and token counters, a list chat:{abc123}:msgs for recent messages with their token counts, and TTLs that refresh on every turn. Each prompt is built newest message first within the input budget, so the oldest turns are the first to be dropped.

This lesson keeps what trimming would lose. It covers rolling summaries of older turns and the structured and semantic memory that lasts across sessions