Mi-Memory: A Lifecycle Memory Framework for Personal AI
This paper introduces Mi-Memory, a lifecycle memory framework for Personal AI that organizes memory management into four roles (Structure, Expansion, Evolution, and Deployment) linked by a shared audit contract to ensure durable, evidence-governed, and deployable continuity across diverse devices.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a super-smart digital assistant that lives in your pocket, your car, and your smart home. Right now, most of these assistants are like goldfish with a very short memory span: they remember what you just said in the conversation, but once the chat ends, they forget everything. They don't really know you—they don't remember that you hate cilantro, that your kid has basketball practice on Tuesdays, or that you once mentioned a fear of spiders.
To make an assistant that feels like a true companion, it needs a "memory." But not just a simple notebook. It needs a memory that can handle photos from your camera, sleep data from your watch, location history from your car, and chat logs from your phone all at once. The big challenge scientists face is how to build a memory system that doesn't just store facts, but understands how they connect, remembers where it got the information, and knows when to forget things or change its mind without making mistakes. This is the world of "Personal AI," where the goal is to create an agent that can grow and adapt with you over years, not just minutes.
This paper introduces a new framework called Mi-Memory (short for "Memory for Personal AI"). Think of Mi-Memory not as a single tool, but as a complete lifecycle for an assistant's brain. The authors argue that memory shouldn't be a static pile of notes; it should be a living process that goes through four distinct stages: Structure (organizing the facts), Expansion (gathering new info from cameras and sensors), Evolution (learning and fixing mistakes), and Deployment (making sure it works on your phone even without the internet).
The team built a system to test this idea, and here is what they found. First, they created a "Structure" layer called MemStack. Imagine this as a librarian who doesn't just throw books on a shelf, but sorts them into tiny facts, daily summaries, and long-term profiles. When they tested this against a specific, controlled reference system called MemBrain (using the exact same testing tools and models to ensure a fair comparison), their librarian was able to keep track of long stories better. On the LoCoMo test, the margin was at a parity-level (very close to the baseline), while on the LongMemEval and PersonaMem-V2 tests, they reported larger relative gains. The authors note that comparing these numbers directly to other public systems is tricky because those other systems often used different models or testing setups, and they explicitly state that MemStack is a controlled-reference anchor rather than a universal leaderboard claim. Within this strict, fair test, MemStack showed its strength in organizing memory into layers to help the assistant remember the right details at the right time.
Next, they tackled "Expansion" with tools called MemSense and MemFuse. This is where the assistant learns to look at the world, not just listen to you. MemSense is like a camera that doesn't just take a picture but tags it with a stable ID, the date, and what was happening in the conversation. MemFuse is like a detective that connects the dots between different devices. For example, if your phone calendar says "basketball practice," your car knows the route, and your home camera sees a sports bag near the door, MemFuse links these three separate clues into one clear story. In their tests, this system was pretty good at connecting these dots, scoring about 89.15% on a visual memory test and showing better ability to figure out cause-and-effect between devices than older methods.
The third part, "Evolution," is about how the system gets smarter without breaking. The authors built a "Governed Evolution" system using D2ACCI and E2MEND. Think of this as a strict quality control team. Instead of letting the AI randomly change its own rules (which could be dangerous), this system uses a "dual-loop" process. A human (or a guided AI) proposes a change, like "let's look at dates more carefully," and then a strict test runs to see if it actually helps. If the test shows the system got worse at anything else, the change is rejected. They found that this careful, step-by-step approach allowed them to fix specific problems—like reducing errors caused by missing information by a huge margin—without accidentally breaking other parts of the system.
Finally, they looked at "Deployment" with LiteMem. This is about making sure the memory works even when you are in a place with no internet or strict privacy rules. They showed that the same memory rules could be translated into simple text files (like Markdown) that live on your device, keeping your data safe and private while still letting the assistant remember things.
The paper doesn't claim to have solved every problem in the world. They admit that some parts, like connecting different devices, are still being tested and are "preliminary." They also note that their system is designed to be "auditable," meaning every time the AI makes a decision, you can trace it back to exactly where the information came from. But the main takeaway is a hopeful one: by treating memory as a full lifecycle rather than just a storage box, we can build personal AIs that are more reliable, more private, and actually capable of remembering who you are.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.