LLMs are inherently stateless; they can't access fresh information, read private data, or take multi-step action on their own. Every agentic technique that exists today is really just a different answer to the same question: how do you give a model exactly the right information, in the right structure, at the right time? Get this wrong, and the failure compounds; a 5% drop in per-step accuracy becomes a 14-point drop in reliability over just three chained agent calls. Get it right, and you've built the difference between a demo and a production system.
In this session, you'll learn how to treat the context window like a suitcase, packing in tools, skills, memories, and session history without letting it become the bottleneck and see how ADK handles this through on-demand tool retrieval, progressive-disclosure skills, and context compaction for long-running agents. We'll go deep on Memory Bank: how raw, noisy session events get extracted and consolidated into refined, durable memories; what gets remembered by default; how to customise it with topic definitions and few-shot examples; and when to inject
Attendees will leave with a practical framework for context control, state persistence, and action routing, plus resources to go deeper on multi-agent systems.