Architecting Agent Memory: Session State, Vector Search, and Managed Cloud Memory

Ask an agent for dinner ideas on Monday, come back Wednesday, and most agents greet you like a stranger. Across eight graded rungs (L0→L7) you'll move from a stateless goldfish → a within-chat scratchpad (session.state) → durable per-user state across restarts (DatabaseSessionService + the user: scope) → searchable long-term memory (add_session_to_memory / search_memory - where persistence is not memory finally clicks) → file artifacts → working-memory management (compaction, rewind, caching) → durability (pause & resume) → and finally the same agent code on managed Vertex services with zero rewrites. One agent, eight rungs - so you always know which one to reach for. 🕒 Americas: Thursday, August 27: 9:00 AM PDT 🕒 Asia-Pacific: Friday, August 28: 9:30 AM IST