Amazon Bedrock AgentCore was a key focus area at AWS re:Invent 2025. Both Matt Garman and Dr Swami Sivasubramanian spent time on agentic AI and AgentCore's central role to productionizing agentic AI. AgentCore exists because AWS prioritizes ecosystems over glueing together different tools. Developers were building cool AI prototypes with LangChain and LLM APIs, but most never got beyond pilots and demos. AWS does not win if nothing gets to production and real usage.
So, AWS set out to solve the problem of AI being hard to scale, risky in prod, too much glue code, no real security or observability. AgentCore is the AWS answer to 'how can customers build their own agents that work and scale reliably?'.
In this blog, we discuss AgentCore and every new enhancement it gets, starting with the newest announcements at AWS re:Invent 2025.
What is AgentCore Policy?
Set agent boundaries using plain language to control which tools agents can use and what actions they’re allowed to perform. AgentCore Gateway enforces these rules at scale. In tests, this handled thousands of requests per second with millisecond policy checks.
AgentCore automatically converts natural language policies into Cedar, an open-source authorization policy language developed by AWS.
My Take
This update fixes one of the biggest blockers to using agents in production. You can now enforce rules like “agents can only access customer data during business hours” without writing any custom auth logic. Since many enterprise POCs get stuck over worries about unrestricted tool access, this pretty much removes that problem.
AgentCore Policy Evaluation
Developers can leverage 13 built-in evaluators that continuously assess agent behavior for correctness, helpfulness, safety, and tool-selection accuracy. The system provides real-time quality scoring with CloudWatch integration, supports custom evaluators, and includes episodic memory to track reasoning patterns and help agents improve over time.
My Take
It removes the burden of creating in-house evaluation or scoring frameworks, offering built-in quality monitoring and CloudWatch alerts by default. The episodic memory capability, which lets agents learn from prior interactions automatically, is a particularly notable improvement.
AgentCore Episodic Memory
"At the heart of what makes agents truly intelligent is memory and context," opined Swami in his re:Invent 2025 day 3 keynote. A new episodic functionality was made generally available to enhance the already existing AgentCore's Memory component. This allows an agent to build a coherent understanding of users over time.
Episodic memory allows an agent to retain and reference past interactions, giving it the ability to recall and reason over prior experiences in a human-like way.
To understand better, let's take a quick detour and go over the different types of memory for Agentic AI. Drawing from cognitive science concepts theoretically, these memory types include -
- Short-term: Working context - immediate inputs, current state of a task or the latest user command. Typically resets when context-shifts or task ends.
- Long-term: Knowledge accumulated over time – patterns, past experiences, or external data it has been trained on.
- Semantic: AI’s “knowledge base” about the world – facts, concepts, relationships.
- Procedural: AI’s “how-to” knowledge – executing a sequence of steps, navigating environment.
- Episodic: Specific experiences or events – agent’s personal history.
AI agents generally operate without true memory. They rely on a short context window that gets wiped every time, so they never learn from what worked (or didn’t) in real-world use.
AWS's Episodic Memory, according to David Richardson (VP AgentCore), is "The idea is to help capture information that a user really would wish the agent remembered when they came back, For example, 'what is their preferred seat on an airplane for family trips?' Or 'what is the sort of price range they're looking for?"
How is episodic memory different?
Episodic memory differs from the earlier AgentCore memory model in that, rather than managing separate short- and long-term memory stores, agents can automatically recall relevant information when triggered by specific conditions. This reduces or removes the need for custom instructions.
How does episodic memory work?
Episodic memory captures meaningful segments of user and system interactions, enabling applications to recall context in a focused and relevant way. Rather than storing every raw event, it identifies important moments, condenses them into compact summaries, and organizes them for efficient, noise-free retrieval.
The episodic memory strategy includes the following steps:
- Extraction – Analyzes in-progress episode, determines if episode is complete.
- Consolidation – Compiles extractions into single episode.
- Reflection – Generate insights across episodes.
Reflections take episodic memory a step further—they look back at past moments to spot patterns and meaningful lessons. Instead of just recalling what happened, they help the system understand why it matters and how it should act next. It’s a way for the model to turn experience into immediate guidance.
Use cases for AgentCore episodic memory
"As hundreds of specialized agents emerged, managing state and maintaining consistent context became increasingly difficult, highlighting the need for a unified memory layer. Amazon Bedrock AgentCore Memory provided the solution through seamless, centralized state checkpointing across our multi-agent orchestration stack. With the new episodic memory functionality, our agents will learn from prior analyses to generate more intelligent insights." said Helene Astier, head of Technology, MI Enterprise Technology and Sustainability, at S&P Global Market Intelligence.
It is ideal in areas and use-cases where events, past interactions add significant value to it. Some of the areas are -
- Customer Support & Service: Agents can remember a user's previous issues, troubleshooting steps attempted, and successful resolutions.
- Personalization: The agent can store user preferences and context across long periods.
- Troubleshooting and Diagnostics: In complex diagnostic flows, the agent can access past "episodes" to identify patterns, common failure modes, and lessons learned from similar situations.
My Take
Episodic memory, in very broad terms, is the agent choosing and creating its own knowledge base. Think of it as its own journal. Along with AWS's launch of Nova 2 Omni and Nova Forge, this functionality reiterates the fact that AWS is moving towards AGI.
This article is part of our comprehensive guide to AWS AI. Explore the guide to know more about AWS AI tools and why AWS AI infrastructure is the best way to build agentic AI solutions that can scale in production.
