AWS reclaims idle memory mid-session in the new AgentCore runtime
Amazon Bedrock AgentCore's new runtime reclaims memory that goes cold during a session and keeps cold start times consistent regardless of container size.
AWS has announced a new version of the runtime layer in Amazon Bedrock AgentCore, the managed compute service used to deploy and run agents. Two things changed: how memory is billed and how sessions start.
In the original runtime, a session held the memory it allocated until the session ended. Usage tracked the high watermark, so an agent that spikes occasionally but sits idle most of the day kept paying for its peak long after it stopped using that memory.
The new runtime starts each session from a small memory profile and pages in additional memory on demand. AWS says it analysed allocation patterns across billions of sessions and tuned the runtime to reclaim memory once it goes cold and is unlikely to be accessed again. The bill follows those changes over the life of the session.
Cold starts
Every new session must start before it can do any work. A session that lands on an already-initialized environment starts in under 100 milliseconds, but guaranteeing that means holding compute in reserve to keep environments hot. So most sessions begin cold: booting a fresh environment, pulling the image, initializing the agent. That penalty grows with image size and concurrency, and is worst under bursty traffic, precisely when the fewest ready environments remain.
According to AWS, the smaller startup profile addresses this directly, delivering consistent cold start times regardless of container size or agent concurrency.
The company notes that customers have been building this machinery themselves: keeping spare environments ready to avoid cold starts, tuning memory allocation, then tearing it down to control cost. AWS describes that as undifferentiated work that has nothing to do with the agent itself.
The full announcement is on the AWS blog.
