Amazon Bedrock AgentCore Reaches GA with Two-Tier AI Memory

The AgentCore platform reached full general availability at AWS Summit New York this week. [1] AgentCore Memory provides AI agents with persistent cross-session memory through a two-tier short-term and long-term architecture. [1] Short-term memory is synchronous and session-scoped, allowing users to call `create_event()` to write raw messages like user inputs and tool calls. [1] This short-term working memory has a configurable retention period ranging from seven days to one year. [1]

Long-term memory is asynchronous and persists cross-session insights indefinitely. [1] An LLM operates in the background to distill short-term events into structured insights, creating an asynchronous delay between storing an event and the insight becoming available. [1] AWS provides three distinct extraction strategies that dictate what the agent retains from raw events. [1] The summary strategy generates rolling conversation summaries organized by topic. [1]

The introduction of AgentCore Memory marks a critical maturation point for AWS's generative AI stack, moving agents beyond stateless, single-session constraints. By splitting memory into synchronous and asynchronous layers, AWS solves the latency problem of real-time recall while still building long-term user context. This compares favorably to early standalone vector database integrations that required developers to build custom extraction and state-management pipelines from scratch. What the sources don't address: The specific cost implications of the background LLM processing required to run the long-term extraction layer.