Personalized Memory Policies Improve Retention for Long-Horizon AI Agents

If an AI assistant remembers yesterday’s trivia but forgets your ongoing project, the real problem is not storage — it is what gets stored. This paper argues that memory systems for large language models should not treat every user the same, because the information worth saving can differ from person to person. To test that idea, the authors introduce PerMemBench, the first benchmark for personalized memory systems, built from multi-year, multi-domain interaction histories across diverse user personas. They also study a lightweight idea called session-level storage gating, which can skip memory operations for transient sessions. The headline result is encouraging: personalization can deliver substantial retention gains when gating is perfect. But the catch is just as important — deciding accurately which sessions deserve memory is still an open challenge. That makes the paper less a finished solution than a clear map of the problem, showing where generic memory policies fall short and what must improve before long-horizon agents can remember the right things for the right people.

A memory system for an AI agent sounds simple until you picture a real user. One person keeps returning to a long project, so yesterday’s detail matters. Another fires off one-off requests, where saving the chat only clutters the vault. That mismatch is the heart of this paper’s surprise: memory does not fail only because it is too small, but because it often stores the wrong things for the wrong person. If your assistant remembers trivia from lunch but forgets the thread you have been building for months, the problem is not just recall. It is judgment — what deserves a place in memory at all.

Why generic memory breaks down

PerMemBench is built around that judgment problem. The benchmark spans multi-year, multi-domain interaction histories and covers diverse user personas, so it can test whether a memory policy still works when people do not behave like carbon copies. The headline finding is careful but clear: personalized policies can deliver substantial retention gains when the system knows which sessions to save, yet the hardest part is not the storage itself. It is deciding, session by session, whether a conversation is transient or worth keeping. In other words, the gains are real, but they depend on a gate that has to guess user intent well enough to avoid wasting the memory budget on the wrong moments.

How the benchmark turns memory into a testable problem

PerMemBench makes personalization measurable by pairing long interaction histories with different user types, then asking memory systems to act under a fixed budget. That setup matters because unlimited storage would hide the issue entirely. The paper also studies session-level storage gating, a light framework that can skip memory operations for sessions that look transient. Used well, the gate keeps short-lived exchanges from crowding out the past that may matter later. Used badly, it throws away the very context an agent needs for long-horizon work. The framework is simple on purpose: the point is to see whether the policy of what to store can be learned in a way that adapts to the user, not just to the model.

Why the memory budget suddenly matters more

This work changes the frame around AI memory. The obvious goal is not simply to remember more, because more memory can also mean more noise. What matters is remembering the right things for the right person, which means the memory system has to behave less like a filing cabinet and more like a careful editor. PerMemBench gives the field a way to test that idea instead of talking about it in the abstract. It also shows why personalization matters so much: if users have different long-horizon needs, a universal policy can spend precious memory on sessions that will never matter again while missing the ones that shape an ongoing task.

What still stands between the idea and a reliable assistant

The next hurdle is not whether personalization can help; the paper already shows that it can. The open question is whether a system can spot, from the session alone, which moments are transient and which ones should be stored for later. That is a harder kind of intelligence than raw recall because it asks the agent to read future value from present context. Until that gate improves, personalized memory will stay partly trapped behind a guessing game. PerMemBench turns that guess into a benchmarkable problem, which is exactly what long-horizon agents need if they are ever to remember like a particular person rather than a generic machine.