I’ve been running into this while adding memory to agents we use across the companies I’m building.
In the paper, structured memory scored 73.6% vs. 44.9% for files. But on abstention, it went the other way: 77.8% vs. 88.9%.
That’s also where I’m struggling most today. Retrieving context is one thing. Knowing what is still valid when information gets old or conflicts is harder.
I’ve been talking to other teams working on this and haven’t seen an approach that handles it really well yet.
I’ve been running into this while adding memory to agents we use across the companies I’m building.
In the paper, structured memory scored 73.6% vs. 44.9% for files. But on abstention, it went the other way: 77.8% vs. 88.9%.
That’s also where I’m struggling most today. Retrieving context is one thing. Knowing what is still valid when information gets old or conflicts is harder.
I’ve been talking to other teams working on this and haven’t seen an approach that handles it really well yet.
How are you dealing with this in production?