Google Cloud's Gemini agent carries one memory into Microsoft 365, Slack and the command line

In Google's own example, a user asks Gemini to set up next week's meeting with the usual group of regional event leads and gives it no names or email addresses ^[9]. The agent works out who they are from a chat space's membership and the thread from the last event. It checks their calendars and starts an email thread to settle a time, even with external participants ^[9].

Every step in that request depends on access someone granted in advance. Google is pitching a teammate that remembers everything it is told ^[3]; what a rollout team signs up for is granting it access, one system at a time.

Thomas Kurian, Google Cloud's chief executive, said: "You give it objectives, not instructions. You delegate an outcome and come back to finished work." ^[4] He also said: "For an agent to do that, it has to be connected to your personal workflows, your systems of record, and your enterprise controls." ^[5]

SiliconANGLE's account names 13 products the agent connects through, among them Confluence, Jira, Salesforce, ServiceNow, Snowflake and Postgres, plus files on desktops ^[8][18]. The design assumes people will not open a new app to hand off work, and the agent needs no dedicated interface ^[2].

It also keeps what it reads. 9to5Google lists four kinds of memory: session, semantic, procedural and episodic ^[11]. Semantic memory is a structured knowledge base the agent builds as it reads documents, talks to people and works with other agents ^[11]. In my view, that changes how a loose file permission behaves. Once the agent has read the file into its knowledge base, the exposure outlives the session. Teams can also publish custom skills to a shared company registry, another store that needs an owner ^[17].

On cost, Google Cloud said per-token prices have fallen about 98% since 2024 while enterprise AI volume has risen sharply ^[14]. One of its answers is saved BigQuery reports that teams rerun without additional token costs ^[15]. Flexible spending options announced in August came online with the launch ^[16]. Neither report includes a price for the agent itself.

For a team weighing a pilot, I'd start with one recurring task, such as a manager emailing for a project update as a slide deck, the case Google uses for its one-click hand-off ^[10]. The test has two axes. The first is whether the task's inputs sit in systems on Google's connector list. The second is whether the permissions on those systems already match who should see the result.

Both yes means pilot, and judge it by whether work comes back finished without rework, the standard Kurian set ^[4], instead of by how often people type @Gemini. Connectable but loosely permissioned means fixing permissions first. That costs weeks before anyone sees a result, and I'd still pay it, because the agent's semantic memory keeps what it reads ^[11]. Inputs outside the list mean integration work before a pilot can start. Neither yes means the task is a poor first candidate.