The Next Memory Wave Won't Be Bigger: It Will Have a Subject

Every few months a new memory system lands with a bigger number: more sessions retained, longer horizons, higher scores on the memory benchmarks. The whole field is running the same race, help the agent remember more. We think the next wave is not about capacity at all. It is about the subject, and this post is the case for why. None of it depends on us being right about who builds it.

Context for new readers: this is part four of our person-centered series. Part one stated the value. Part two reviewed nine memory systems and found the same structural gap in all of them. Part three was the audit of our own data, where 46% of personal content could not be resolved to a specific person. This entry stands alone, but the series is the argument.

Pressure one: the scoreboard is changing

Benchmarks decide what gets built. The first wave of agent memory was measured on factual recall over long horizons: LoCoMo (2024) with dialogues spanning up to 35 sessions, LongMemEval (2024) with 500 questions across five abilities. Both ask one thing: did the agent remember the fact? The field optimized for exactly that, and got good at it.

The next question is already on the scoreboard: did it remember the person? PersonaMem (2025) builds 180+ simulated user histories of up to 60 sessions each, then tests whether a model can track how a user's situation evolves over time. Frontier models (GPT-4.1, o4-mini, GPT-4.5, o1, Gemini-2.0) land around 50% overall accuracy. The revealing detail is not the headline number, it is the shape of the errors: models answer questions about static facts far better than questions that require accounting for the user's latest situation. Remembering is close to solved. Remembering who is changing is not.

Read the December 2025 survey of the field and the same story shows up in the taxonomy. It organizes agent memory by form (token-level, parametric, latent), function (factual, experiential, working), and dynamics (formation, evolution, retrieval). It is a useful map, and we use it ourselves. But notice what is not an axis: the subject. A benchmark that measures the subject, inside a taxonomy that lacks one, is how a research direction gets born. What gets measured gets built.

Pressure two: "remembers everything" has a failure mode

Assume the capacity race succeeds. The agent remembers everything you ever said. Now watch it fail anyway.

It recites your details back perfectly and mixes up the people in your life. It knows you mentioned your sister in March and cheerfully suggests a catch-up call in June, after the funeral. It tells your sister what you told it about your job, because "user's family" was one bucket in the store. None of these are retrieval failures. Every fact is in there. The failure is that nothing in the system knows which person each fact belongs to, so nothing can scope an answer to the right one.

Recall without a subject is uncanny before it is useful. Users do not experience it as intelligence; they experience it as a stranger reading their diary with confidence. And as context windows grow, capacity stops differentiating anything at all. Attention is the scarce resource now. For the queries users actually care about (who, about whom, what changed), retrieval that routes through a person graph beats retrieval by text similarity, because similarity cannot tell your sister from your coworker. Both of them used the word "deadline".

Pressure three: sovereignty is becoming a requirement, not a feature

"Your data belongs to you" is a promise every privacy policy makes and almost no memory system can keep. Look at what the promise actually requires: show me everything you know about me; show me where each piece came from; let me delete one piece without collateral damage; let me take all of it with me when I leave.

An account-rooted schema cannot do any of that cleanly. Delete the account and you leave a person-shaped hole in a graph that still references them. Export the account and you export a bag of facts with no subject attached to any of them. The regulations keep moving toward exactly these rights, and the technical precondition for honoring them is a person root in the schema, not a nicer privacy page. Portability is the same story: a memory export that cannot say whose memory it is, is not an export of your life. It is a dump.

Pressure four: the second person is already in the house

Every assistant in production today serves one account, and the entire memory layer is built on that assumption. That assumption is breaking now.

Shared devices are normal. Family assistants are shipping. Team agents are a category. Home robots are the obvious endpoint, and they arrive into a house with several people in it. The moment a second human appears, single-account memory fails in a specific way: it cannot tell two people apart, cannot scope a memory to one of them, and will eventually tell your guest what you told it about your family. Nobody ships a multi-person memory layer on purpose. It gets built when the second person shows up and the first workaround stops working, which is the worst possible time to design a schema.

What would make us wrong

Honest version: we could be wrong about who builds this. A lab with a foundation model and a distribution channel could ship a person-centered memory layer as a feature and skip us entirely. We do not think we are wrong about whether. The measurement is moving, the failure mode of pure capacity is visible to anyone who has watched an assistant confuse the people in their life, the sovereignty requirements are hardening, and the second person is already in the house. Somebody will build the subject layer. We would rather it be built by people who think the subject is the point.

Which is the value, one more time: everyone is unique, and a generic memory cannot solve everyone's problems. Not because the models are weak, but because the storage was built for nobody in particular.

References

1. Maharana, A., Lee, D.-H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024). Evaluating Very Long-Term Conversational Memory of LLM Agents. arXiv:2402.17753. ACL 2024.
2. Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., & Yu, D. (2024). LongMemEval: Benchmarking Chat Assistants with Long-Term Interactive Memory. arXiv:2410.10813.
3. Jiang, B., Hao, Z., Cho, Y.-M., Li, B., Yuan, Y., Chen, S., Ungar, L., Taylor, C. J., & Roth, D. (2025). Know Me, Respond to Me: Benchmarking LLMs for Dynamic User Profiling and Personalized Responses at Scale. arXiv:2504.14225.
4. Memory in the Age of AI Agents (2025). Survey. arXiv:2512.13564.