Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
A new research paper introduces NAPM Mem, a framework that transforms long-term user memory from passive retrieval into a structured, multi-granularity action space for AI agents. By organizing memories into a pyramid of raw conversations, typed records, topic tracks, and user profiles, agents can actively navigate and explore memory using tools, dramatically improving accuracy and reducing noise. The presenter demonstrates this live, showing an agent retrieving specific past events and conversations by traversing multiple memory layers.
Key points
- Traditional AI second brains break down as information accumulates because retrieval becomes noisy and confused.
- The NAPM Mem framework treats long-term user memory as a structured action space rather than passive retrieval context.
- Memory is organized into a multi-granularity pyramid: raw conversations, typed memory records, topic tracks, and user profiles.
- All memory levels are connected through provenance relations and exposed to agents as tools for active exploration.
- The approach couples structured storage (like Convex) with a learned policy for using memory at the appropriate granularity.
- A live demo shows an agent navigating multiple memory layers to retrieve specific past events and conversations.
- Accuracy improved dramatically, from 50% stealthness to 100%, after implementing the framework.
- The framework supports multiple user identities within the same action space.
Tools mentioned
Techniques
- Multi-granularity memory pyramid
- Structured action space for memory
- Provenance relations between memory levels
- Learned policy for memory granularity selection
- Active memory navigation with tools
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
I don't remember when I was this excited after reading a research paper than this one. From passive retrieval to active memory navigation, learning to use memory as a structured action space. If you ever wondered, then you send a message to your agents or AI second brain where does it go? How does it work? And if you use the second brain in agent memory, you know, the more information and the more of this noise there [music] is, retrieval breaks, it gets confused.
This research paper fixes it because instead of retrieving some information using semantics or time codes or chunking or whatever it is, we give [music] agents tools to explore our memories. And same as humans, we have a layers of our thinking, critical, analytical, creative, abstract, concrete, but we use different kind of methodologies and tools to unlock those different ways of thinking. And AI, they always said it's a thinking partner. Use it to enhance your thinking, not to replace it. We don't build AI second brain.
So we don't think and don't remember anything and just ask it and it gives us information. We use it so that we can actually think bigger, [music] expand our thinking. I worked with multi-agent systems. I built the open claw secure and safe replica which I've been using for the last months and my community tested that one of the best agents they still use to this day. I had to implement everything what this paper said.
So Quen business unit introduced this paper. Okay, it is July 2026. They introduce NAPM mem a framework for learning to use longterm user memory as a structured action space rather than passive retrieval context. NAPM organizes user history into linked multi-granularity memory pyramid where raw conversations, typed memory records, topic tracks and user profiles. [music] So it could also be not just one identity but maybe multiple users using the same action space.
All this is connected through provenence relations and exposes these levels through memory tools. Our results suggest that long-term user memory benefits from coupling structured storage like what we using goat like convex [music] with learned policy for using memory at the appropriate granularity. If you do anything today, take this research paper, toss it [music] in your AI. If it's Hermes, if it's open claw, if it is clot code, doesn't matter. Come to the community, check out the gobot, set it up.
And the last thing I want to show you is what happens if we send live message because I wanted to see how is this agent actually using tools to navigate my whole memory. I will test two things about something that happened months ago and something that happened last week. What was my um call about together the short end of the week sync call? Can you tell me from your memory? I'm being specific from memory because the easiest is to go to granola and test the meeting transcripts but I wanted to go to the memory.
So let's see what is going to happen. I send it a voice message but I don't see anything happening. Ooh, look. Okay. Voice message, transcript.
You're literally seeing everything on the screen at the same time. At the same time, it's searching conversations. Okay. Tool search. It's so Wow.
It's going into messages. Wow. Okay. I'm not touching anything. I want to kind of zoom in, but you can see all the links and connections that it's exploring.
It's going to go about phone call. Wait, if I touch it, I want to see from different layer. It's going through topics, clusters. It went to granola. It went to the gobot calls.
It went through the different facts. Found shorts. found got a weekly agenda meeting. I can see all the tools that it used. And if anyone is now, hey, what's about the latency?
You just put a faster model behind it. Now it's not using the fastest model. I said, what did I discuss on my latest call? We short last week. Last week and short.
So it had to go through multiple layers of reasoning. So I had today's call, but we talked about memory heist video. Yes. Yep. [music] And this is the sync call that we had with sword.
It's crazy. The last thing I'm going to ask is about Tenerifa. Hey, when was I planning to go to Tenerifa? I'm not saying here to search my memory. Okay.
So, I just send voice message. Let's see what happens. [music] Bam. At the same time, it's searching and at the same transcribing. got three hits.
Gobot memory activated. The the great thing here is that we humans live in a paradox. Two things can be true at the same time. Sometimes in AI systems memories, you want to surface verbatim messages. In this one, you can literally fetch exactly the text as well.
Okay. The trip already happened. It ended roughly March to May and it wrapped up in May. You've been back in Berlin since. So these are clusters depending on the topic.
As you can see I have privacy because this is literally my bread and butter. This is my involve information and I know that some of them are just like a business some of them are health and you can see that those messages many of those are just like noise. Then we have records and facts and traces and those are not one kind of like a graphbased system what you would see in Obsidian or in a lot of AI second brains. It is not just like a pieces of files or chunks of information. You have hierarchy and [music] structure and it works like a pyramid.
And yes, I ran ewells and this went from stealthness 50% to 100%. It's insane. [music] Accuracy improved everything. Our accelerator is already updated. We were teaching rag and semantic search teaching [music] actively exploring memory of your agents.
This is the new thing and it's materials are all updated. Gobot is updated. If you are in my community, please pull this update. It is [music] really much better than before. And if we are meeting first time, please check out our community.
We have a free trial. Join our 8week AI accelerator and come join us on cloud certification. You're going to find the research paper link down below. And right now, YouTube is going to recommend you a video that you might like. So, let me know in the comments below if YouTube was right.
Thank you and see you in the next