How do you give AI agents persistent memory without bloating your context window and spiking latency? Developer Relations Engineer Annie Wang explains how to architect two-tier agent memory:
* Short-term memory (SessionService in ADK): Tracks active turn history, tool outputs, and working scratchpad state within a single task.
* Long-term memory (MemoryService in ADK): Extracts durable user preferences and project conventions across sessions, indexing them for selective retrieval.
* Context compaction: Using rolling summarization and pruning so multi-hour agent workflows stay fast, accurate, and cost-effective.
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
#GoogleCloud #AIAgents #CloudArchitecture
Speaker: Annie Wang
Products Mentioned: Gemini Enterprise Agent Platform
* Short-term memory (SessionService in ADK): Tracks active turn history, tool outputs, and working scratchpad state within a single task.
* Long-term memory (MemoryService in ADK): Extracts durable user preferences and project conventions across sessions, indexing them for selective retrieval.
* Context compaction: Using rolling summarization and pruning so multi-hour agent workflows stay fast, accurate, and cost-effective.
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
#GoogleCloud #AIAgents #CloudArchitecture
Speaker: Annie Wang
Products Mentioned: Gemini Enterprise Agent Platform