EvoMemState: Evolving Memory State for Personal Agents

EvoMemState. Evolving Memory State for Personal Agents. By Zhishang Xiang and nine others, from Xiamen University, Nanyang Technological University, and three other institutions. Published on the second of November.

A personal agent built on a large language model is only as useful over time as what it remembers. Long-term memory lets it retain user information across interactions and keep its assistance continuous and personalised. Existing memory systems handle this through a pipeline of fact extraction, organisation, updating and retrieval. Yet they largely keep memories as independent facts, each one standing alone. The authors argue that this fact-centric paradigm can miss how a new fact updates an earlier one, constrains it, or conflicts with it. It also leaves implicit whether a fact applies to the request at hand, what role it plays there, and how much priority it deserves.

EvoMemState shifts the unit of memory from facts to states. Persistent facts are kept, but they are augmented with two complementary kinds of state. Relational States capture how facts remain valid and how they relate to one another as memory evolves over time. A Reasoning State, by contrast, belongs to the current request: it determines how those facts should be applied to what the user is asking right now. One kind of state looks across the history, the other looks at the present moment.

Three components carry this out. The first, Ontology-Guided Memory State Induction, extracts Persistent Facts and builds the Relational States that maintain each fact's validity, its scope, and its evolving relations to the rest. The second, RL-Based State Evolution, uses reinforcement learning to internalise experience of how memory gets used into a policy, and that policy generates a Reasoning State specific to each request. The third, State-Aware Reasoning, brings the pieces together. It combines the relevant facts with their Relational States and the freshly generated Reasoning State, and uses that combination to guide the personal agent's answer.

The evaluation runs on long-term dialogue benchmarks and on memory-use benchmarks. Across these, the authors report consistent improvements over existing memory systems on two measures: retrieval completeness, meaning how fully the relevant memories are recovered, and memory-grounded response accuracy, meaning whether the agent's replies are correct given what it should remember. The abstract does not give the individual figures, only that the gains hold consistently.

Underneath the three components sits one reframing. A memory store that holds facts as isolated entries treats remembering as lookup. EvoMemState treats it as bookkeeping on relationships plus a judgment made fresh for each request: which facts still hold, how they bear on one another, and how they should shape this particular answer. The reported gains in completeness and grounded accuracy are the evidence offered that this extra structure pays for itself.