arxiv:2611.03743

EvoMemState: Evolving Memory State for Personal Agents

Published on Nov 2

· Submitted by

eric on Nov 6

Authors:

Abstract

Long-term memory enables large language model (LLM) agents to provide continuous personalized assistance by retaining user information across interactions. Existing systems support this process through fact extraction, organization, updating, and retrieval, but largely maintain memories as independent facts. This fact-centric paradigm may overlook how new facts update, constrain, or conflict with earlier ones, while leaving their applicability, role, and priority for the current request implicit. In this paper, we introduce EvoMemState, a state-centric framework that augments persistent facts with two complementary states: Relational States that capture how facts remain valid and relate to one another as memory evolves, and a Reasoning State that determines how these facts should be applied to the current request. Specifically, EvoMemState consists of three key components: (i) Ontology-Guided Memory State Induction extracts Persistent Facts and constructs Relational States that maintain their validity, scope, and evolving relations. (ii) RL-Based State Evolution internalizes memory-use experience into a policy through reinforcement learning to generate a request-specific Reasoning State. (iii) State-Aware Reasoning then combines the relevant facts with their Relational States and the generated Reasoning State to guide the personal agent. Extensive experiments on long-term dialogue and memory-use benchmarks demonstrate consistent improvements over existing memory systems in retrieval completeness and memory-grounded response accuracy. Our code and implementation details are available at https://github.com/XMUDeepLIT/EvoMemState.git.

View arXiv page View PDF GitHub 17 Add to collection

Community

eric-xiang

Paper submitter about 13 hours ago

Long-term memory enables large language model (LLM) agents to provide continuous personalized assistance by retaining user information across interactions. Existing systems support this process through fact extraction, organization, updating, and retrieval, but largely maintain memories as independent facts. This fact-centric paradigm may overlook how new facts update, constrain, or conflict with earlier ones, while leaving their applicability, role, and priority for the current request implicit. In this paper, we introduce EvoMemState, a state-centric framework that augments persistent facts with two complementary states: Relational States that capture how facts remain valid and relate to one another as memory evolves, and a Reasoning State that determines how these facts should be applied to the current request. Specifically, EVOMEMSTATE consists of three key components: (i) Ontology-Guided Memory State Induction extracts Persistent Facts and constructs Relational States that maintain their validity, scope, and evolving relations. (ii) RL-Based State Evolution internalizes memory-use experience into a policy through reinforcement learning to generate a request-specific Reasoning State. (iii) State-Aware Reasoning then combines the relevant facts with their Relational States and the generated Reasoning State to guide the personal agent. Extensive experiments on long-term dialogue and memory-use benchmarks demonstrate consistent improvements over existing memory systems in retrieval completeness and memory-grounded response accuracy. Our code and implementation details are available at https://github.com/XMUDeepLIT/EvoMemState.git.

orinmemory

about 11 hours ago

Treating memory as evolving Relational plus Reasoning states instead of a bag of facts is a useful shift for personal agents. On MemSyco-Bench with Qwen3-8B, EvoMemState reaches 57.6% accuracy, about +12 points over the strongest baseline, with the lowest Syco. Rate and highest Correct Use in that setting. I like that the Memory State Evolution Policy is a separate Qwen3-4B trained with GRPO, so the answer model is not inventing its own memory-use policy mid-response. One practical question: when user preferences reverse several times in a short window, does the Reasoning State mostly gate stale facts, or do you still need an explicit conflict-resolution rule in the ontology?

tangyunbo

about 5 hours ago

Thank you for your interest in our work and for your thoughtful question! When a user reverses their preferences several times within a short window, the ontology and Reasoning State need to work together, because neither is sufficient on its own.

With the ontology alone, we can explicitly record updates and unresolved conflicts between preferences, but we still need to determine what those relationships mean for the current request. For example, an older preference may no longer apply when making a recommendation, yet remain important evidence when explaining how the user’s preferences have changed. The ontology provides the relational grounding, while Reasoning State determines each memory’s applicability, role, and priority for the current request.

Conversely, without maintaining these relationships in the ontology, Reasoning State would need to reconstruct the preference history from scattered facts for every request. With frequent reversals, this increases the risk of overlooking contextual conditions, confusing an update with a conflict, or making inconsistent judgments across requests.

Therefore, we still need explicit update and conflict relationships in the ontology. Reasoning State does more than gate stale facts: it determines how those relationships and the associated facts should guide the current response. Together, the two layers help distinguish preferences that have genuinely been superseded, preferences that apply in different contexts, and conflicts that remain unresolved, rather than applying a blanket “latest statement wins” rule.