Why it matters
Long conversations are common. Compression enables them. Understanding shapes long-conversation apps.
Advertisement
The architecture
Recent turns: full text.
Older turns: summarized by LLM.
Very old: dropped or archived.
Advertisement
How it works end to end
Sliding window: fixed N recent turns; drop older.
Progressive summarization: summarize when approaching limit.
Semantic memory: extract facts to long-term store.
Trade-off: freshness vs context.