What Makes Good AI Agent Memory?

This post is adapted from something I originally shared on LinkedIn. It was prompted by an interesting weekend read from the Wall Street Journal that got me thinking more about AI agent memory, a topic I’ve also been thinking about a lot in my work recently.

Agents are increasingly moving into more of our personal and professional workflows. And for agents, context is everything!

Without the right context, agents can produce responses that are incorrect, irrelevant, or just not useful. Once that happens a few times, users lose trust.

But there’s an important nuance here. As context windows get larger, it is tempting to think we can just give agents more information and let them figure it out. But more context is not always better context.

The challenge is giving agents the right context at the right time. From what I’ve seen and heard firsthand, one of the biggest challenges for agent builders is long-term memory. Not just storing facts, but also learning and storing memory that agents can use in a way that is helpful to users. That raises a question I’m still thinking through: what makes agent memory “good”?

To me, good agent memory needs to be:

• Accurate
• Timely
• Relevant
• Secure
• User-controlled

I find the analogy to machine learning useful here. In traditional ML, model quality is only as good as the features (data) used to train and serve the model. Similarly, an agent’s usefulness across multiple sessions is only as good as the quality of the memories and context it can access.

The article does a great job highlighting what can happen when memory goes wrong. A memory that is stale, overly personal, or used in the wrong context can make an agent seem less helpful, and sometimes even uncomfortable. Building agent memory is not just an exciting technical challenge. It is also a real responsibility we have as agent builders.

Curious how others are thinking about this: what do you think defines good AI agent memory?