Step 3 · Remember

How it works

Sessions contain the events and state of one conversation. Agent Runtime stores them for you, and the agent uses VertexAiSessionService. A new session starts empty.

Memory Bank contains facts about a user, for each app and user id. After each turn, a plugin sends the turn to Memory Bank. Memory Bank uses an LLM to extract facts on the topics that you allow. Then it merges these facts with the facts that it has. At the start of the next turn, PreloadMemoryTool searches the facts with the new message. It adds the matches to the instructions of the model.

1class MemoryBankPlugin(BasePlugin):2    async def after_run_callback(self, *, invocation_context):3        events = [e for e in session.events if e.invocation_id == invocation_context.invocation_id]4        await memory.add_events_to_memory(..., events=events,5            custom_metadata={"wait_for_completion": False})

You choose the facts that Memory Bank keeps. infra/memory/memory-bank.json limits the extraction to two topics, team_and_office and work_preferences, and to explicit instructions. In step 2, grant_agent_access.sh applied this file.