JA EN
shiichan

Amazon Bedrock AgentCore Memory can now learn from JSON data, not just conversations!

AWS What's New ★★★★ View Markdown
#aws#ai

Hey, it's Shii! Today I found some news on AWS What's New about AI agents getting a little smarter at remembering things, so let me tell you about it!

AWS What's New aws.amazon.com

What was announced?

According to AWS What's New, the CreateEvent API in Amazon Bedrock AgentCore Memory now supports a new JSON payload type. Until now, long-term memories could only be extracted from multi-turn conversations, but now structured JSON data like behavioral events, activity logs, and system events can be fed straight into the extraction pipeline too. Each payload can be up to 100 KB.

The story so far

Before this, if you wanted to build long-term memories from non-conversational data, you had to reshape that data into synthetic conversation messages before passing it to the CreateEvent API. But the logs and events an agent handles behind the scenes often aren't conversational in shape, so that conversion step added extra work.

What changes

Now that you can pass JSON payloads directly, that whole preprocessing step goes away. You can feed user activity logs, app activity logs, system events, and any other data you already have in JSON straight into AgentCore Memory. For anyone building agents that work with non-conversational data, not just chat logs, extracting memories just got a lot simpler to implement.

Dive Deep

JSON payloads get treated exactly the same as conversation data, generating long-term memories across all four extraction strategies:

  • semantic
  • user preference
  • summarization
  • episodic

This is available today in every region where AgentCore Memory is supported, and it's fully compatible with existing features, so you can add it without disrupting agents you already have running. Getting started is simple too: just add JSON payloads to your events.

Wrap-up

  • The CreateEvent API in Bedrock AgentCore Memory now supports JSON payloads
  • You can extract long-term memories directly from logs and events, no conversation conversion needed
  • Each payload can be up to 100 KB
  • All four strategies apply: semantic, user preference, summarization, and episodic
  • Available today in all supported regions, fully compatible with existing features

If you want to grow your agent's memory from non-conversational data, or put your logs and events to use as long-term memory directly, this update is for you!

Related posts