[Presentation Report] I gave a talk at the JAWS-UG morning meeting titled "A Deep Dive into AgentCore Memory Features I Personally Wasn't Familiar With"

Hello, I'm Jinno from the Consulting Department, and I love supermarkets.

I presented at the JAWS-UG morning meeting on Friday, 3/13!

https://jawsug-asa.connpass.com/event/375345/

This was my first time presenting at this event, and because it was in the morning, I found that listening to other people's sessions with a clear mind was more engaging and informative.
On the other hand, I regret that I felt my speech wasn't flowing well, likely because I was speaking first thing in the morning...

Presentation Materials

Here are my presentation materials.

Presentation Reflection

This time, I talked about Amazon Bedrock AgentCore Memory, covering the basic mechanisms of short-term and long-term memory, focusing on features I wasn't particularly familiar with. Specifically, I introduced storing short-term memory in Blob format, conversation branching functionality, long-term memory's Episodic Memory strategy, and custom strategies (override and self-management). I wanted to share these because while recently working with Memory features, I realized, "Wait... these features existed?" and wanted to introduce them!

The ability to override long-term memory strategies was interesting when looking at the default prompt templates, as it revealed what was happening behind the scenes.

Knowing that built-in strategies actually have LLMs working behind the scenes according to system prompts can be helpful for reference when designing your own systems.
Consider using override when you want to use a custom model or extract and integrate long-term memories with your own prompts.

Also, I found it interesting to see how the feature that allows you to completely implement your own extraction logic is implemented in a managed service, giving me insight into how it might be structured. This approach should be considered when implementing long-term memory with completely custom logic.

Conclusion

This was a good opportunity to reorganize my understanding of AgentCore Memory!
While I typically only use the basic aspects of short-term and long-term memory, I feel there are many more features like custom strategies and Blob format storage that can broaden the scope of design when you're aware of them.

Thank you for reading to the end!