AWS AgentCore brings Memory to GovCloud

Amazon Bedrock AgentCore gains memory, a policy layer governing tool access and managed agent startup in US GovCloud West. Identity and observability were already there. The three new pieces are what make agents in the regulated zone operable in the first place.

Key takeaways

Related:AWS Identity Center: Multi-Region is now part of the setup  /  AWS Cognito: Skill resets the client

Three pieces finally make the partition complete

What is Amazon Bedrock AgentCore? AgentCore is the AWS platform for operating AI agents. It provides Runtime, a Gateway to tools and data, Identity, observability and now Memory plus Policy. Framework and model remain selectable. The service takes over the infrastructure work that every team has built for itself so far.

AWS added the three missing operational pieces in GovCloud US-West on 7 August 2026. Memory stores the session and longer-term preferences. Policy decides on tool access outside the agent code. The managed Harness starts an agent from configuration, without teams having to build their own orchestration loop.

For teams with US government customers, this is the difference between a prototype in the commercial cloud and an agent in the regulated partition. For everyone else, the news is a clear picture of the supported feature set: AWS now specifies which agent components are available in a strictly isolated partition and which are not.

Memory separates session and long-term memory

AgentCore Memory has two layers. The short-term layer retains the ongoing conversation so the agent can continue working on the same task in its next response. The long-term layer retains insights and preferences across sessions. AWS promises personalized flows without teams having to run their own storage infrastructure.

The Harness makes this separation tangible. Every session runs statefully in an isolated micro-VM. Files and memories can outlive the lifetime of that VM. When the underlying session ends, the memory stays reachable as soon as Memory is attached.

That is exactly why Memory belongs in the inventory before the first production agent starts in GovCloud. Without long-term memory, the agent starts from scratch whenever a user returns. With long-term memory, content that nobody treats as transient chat moves into a persisted store. Decisions about data storage determine which fields end up there. The model does not determine this.

Policy checks the tool call before the code

Policy in AgentCore sits centrally in front of the tools. Teams write rules in natural language. AWS translates them into Cedar, the open policy language from AWS. The rules are attached to the Gateway. Every request to a tool is allowed or rejected there before the agent executes it.

The benefit is decoupled from the prompt. Teams that keep the rule in the Gateway can change it without retuning the model. Teams that put it only in the instructions remain vulnerable to prompt drift. In a regulated partition, this is the real operational gain: control sits alongside the agent, outside its text.

The limitations are immediately apparent. Bedrock Guardrails Policy and Temporal Policy are missing in GovCloud. The CloudFormation resources for Policy and Policy Engine are missing as well. Teams managing infrastructure as code must check the policy configuration process in the console and through the APIs. AWS does not provide full infrastructure-as-code coverage of the commercial feature set in this partition.

The Harness turns configuration into an agent

Every agent needs a loop: call the model, pick a tool, return the result, keep the context, catch errors. The managed Harness turns that into a configuration. Model, tools and instructions are declared. AgentCore provides the environment, compute capacity, storage, identity and observability.

AWS specifies several concrete characteristics. The session is stateful by default. It runs in an isolated micro-VM provided by the Runtime. The agent may use the file system and the shell. Models come via Amazon Bedrock, OpenAI, Google Gemini or a LiteLLM-compatible provider. Switching mid-session is documented, and the context is preserved.

Tools attach to the Gateway, to MCP servers or to the built-in browser and code interpreter tools. Custom containers are possible when the standard environment is not enough. There is no separate Harness price. Charges are based on the AgentCore components used.

The Harness therefore saves weeks of infrastructure work. It does not replace an architecture decision. Whoever enables the browser or Code Interpreter with internet access is deliberately sending data to destinations outside the partition. The GovCloud documentation says this clearly: with these customer configurations, data leaves the region via the path configured by the team.

Registry, web search and instances stay out

The feature matrix for GovCloud US-West is the real checklist. In: the Harness, Runtime micro-VMs, Memory, Gateway, Identity, built-in tools, observability, Policy and Evaluations. Out: Runtime instances, Payments, Optimization, AWS Agent Registry and the Web Search tool.

The GovCloud differences list adds further gaps. The Gateway has no semantic search. Agent Registry, currently in preview, remains unavailable. Guardrails Policy and Temporal Policy are missing. Six CloudFormation resources do not exist, including Evaluator, the OAuth2 and API key providers, and Policy and Policy Engine.

Metadata such as runtime names, Gateway targets or evaluation labels must not carry export-controlled content. AWS may only pull such metadata out of GovCloud when the customer opens a support case. Whoever runs agents in this partition keeps names and tags deliberately lean.

The gaps list applies to Frankfurt too

Memory, Policy and the Harness are already available in the Frankfurt commercial region. The 7 August announcement therefore does not mark an EU launch. It shows the most restrictive publicly documented feature subset that AWS currently makes available for AgentCore.

Teams planning to run agents in a European sovereign cloud environment or an environment with strict C5 requirements can use the GovCloud list as a test case. If the Registry is missing, if policy as code is missing, if web search is missing, that is no peripheral detail. What is missing then is the operational catalog, the repeatable policy path or the regional search.

The sensible first step is an inventory against exactly this matrix: attach Memory or leave it out deliberately, Policy at the Gateway, independent of the prompt, the Harness only with clearly set browser and interpreter targets, and name the CloudFormation gaps in advance. The announcement makes agents in GovCloud possible. It does not make them complete.

Source: AWS announcement of 7 August 2026, the AgentCore release notes, the GovCloud differences list and the regional feature matrix.

Frequently asked questions

Image source: AI-generated (August 2026)

Translated from the German original using artificial intelligence. The German version is authoritative.