Gemini Memory Bank GA: How Google's Agent Platform Memory Bank Works, Sessions & Global Routing in June 2026

The question that determines your architecture changed yesterday. Gemini Enterprise Agent Platform Memory Bank and Sessions — the two stateful primitives at the core of Google's managed agent runtime — moved to General Availability on June 17, 2026, with a capability that was not there in Preview: global and multi-regional endpoint support.

The GA announcement is brief in the release notes. The operational implication is not.

What happened

Google promoted Memory Bank and Sessions multi-regional and global endpoint support from Preview to GA as part of the Gemini Enterprise Agent Platform. The short entry in the release notes carries a hard constraint: Customer-Managed Encryption Keys (CMEK) cannot be used if your Memory Bank or Sessions instance is configured to use the global endpoint.

That single sentence will stop several enterprise deployments cold. Teams that have been planning for Memory Bank GA while assuming they could combine global routing convenience with their CMEK posture will need to redesign.

Memory Bank is Agent Platform's cross-session persistent memory layer. It stores structured facts about users and contexts, enables agents to recall past interactions, and supports multi-turn continuity without engineering a custom memory store. Sessions is the runtime session-tracking primitive that ties together agent requests across multi-turn conversations. Both are now GA, which means Google considers them production-ready and they fall under standard SLAs.

Why it matters for AI engineering teams

Stateful agents have a fundamentally different routing profile than stateless inference calls. A stateless call goes to any available endpoint; a stateful agent call must maintain session affinity or tolerate a structured handoff. Memory Bank GA changes four decisions simultaneously for teams on Google's platform:

1. Regional endpoint vs. global endpoint: The global endpoint simplifies routing — you point at one URL and Google handles regional distribution. But CMEK doesn't work with it. If your compliance posture requires customer-managed keys (HIPAA, FedRAMP, or enterprise data governance policies), you are forced to pick a specific regional endpoint, which means you own the failover logic when that region degrades.

2. Multi-regional endpoints: The supported regions now include US multi-region, EU multi-region, and individual zones across Asia-Pacific including Tokyo (asia-northeast1), Seoul (asia-northeast3), Hong Kong (asia-east2), and Mumbai (asia-south1). If your users are across multiple continents, a multi-regional endpoint balances latency without requiring single-region commitment — but again, CMEK is incompatible.

3. Memory Bank default model: The docs note that Memory Bank uses gemini-2.5-flash as its default model for memory summarization and retrieval. For regions where regional Gemini availability is not present, the global endpoint is used automatically as a fallback. This is an invisible behavior change: teams with regional compliance constraints that are also operating in under-covered regions may inadvertently route memory operations through the global endpoint, bypassing CMEK even if their primary inference endpoint is regional.

4. Sessions and Memory Bank share the same endpoint locality rules: A session object created in us-central1 must be continued there. If your agent routing tier is sending requests across regions for load balancing, any stateful call that tries to resume a cross-region session will fail. This is not new in GA, but GA's promotion removes the "this is Preview behavior, subject to change" escape hatch.

The router/operator angle

For teams using an AI gateway or router in front of Gemini Enterprise Agent Platform, Memory Bank GA introduces a new class of request that your routing layer must treat differently.

Session-bearing requests must be routed to the region where the session was created. If your gateway is currently routing Gemini API calls for load balancing or failover without tracking session locality, Memory Bank GA is the forcing function to add session-affinity awareness to your routing policy. The cleanest pattern: attach the session region as a request header or parameter at session creation, and have your router extract and pin it for the lifetime of that session.

Memory Bank calls with CMEK requirements must be explicitly routed to regional endpoints. If your gateway is currently forwarding to the global endpoint by default, add a request classification layer: calls where CMEK is a policy requirement must use a regional endpoint path. Requests that can tolerate no-CMEK can use the global endpoint for latency benefits.

The implicit global-endpoint fallback for under-supported regions is a compliance risk. If your Memory Bank instance is configured for a regional endpoint but a memory operation hits a region where Gemini 2.5 Flash is not regionally available, the platform falls back to the global endpoint silently. Your router or monitoring layer needs to detect this: watch for endpoint_type: global in response headers when you expect regional, and alert.

What TheRouter users should watch or try

Memory Bank GA is specific to Gemini Enterprise Agent Platform. If you are routing Gemini inference calls through TheRouter using the standard /v1/chat/completions path, Memory Bank GA does not directly affect your inference routing today — Memory Bank is a managed, platform-side primitive, not an inference parameter.

However, if you are evaluating whether to build stateful agent pipelines on Google's platform vs. maintaining your own memory layer and routing inference through TheRouter:

• Prefer TheRouter + your own memory store when CMEK is required and you need global routing flexibility. Combining a custom memory store with TheRouter's provider-independent routing gives you both CMEK control and the ability to fall back to alternative providers.
• Gemini Enterprise Agent Platform's Memory Bank GA makes sense when you are already committed to the Google platform, CMEK requirements are absent or you accept regional-only endpoints, and session affinity is manageable within a single-region deployment.
• Watch the Memory Bank default model: gemini-2.5-flash is used for summarization. If you later route inference to a different Gemini tier through TheRouter for cost reasons, your memory-layer model and your inference model may diverge, affecting recall quality.

The full supported locations list — including the CMEK × global endpoint constraint — is documented at Supported locations for agents.