EvoMemState: Evolving Memory State for Personal Agents

| Source: arXiv AI

Tags: EvoMemState, agent memory, personal agents, reinforcement learning, long-term dialogue

EvoMemState gives LLM personal agents memory that tracks how facts relate, conflict and apply to a request, using RL to build a reasoning state, and reports gains on long-term dialogue benchmarks.

Details

Long-term memory systems for LLM agents usually store facts independently, missing how new facts update or conflict with old ones and how they should apply to the current request. EvoMemState adds two kinds of state to persistent facts. Relational States, built through ontology-guided induction, track each fact's validity, scope and evolving relations. A Reasoning State, generated by a policy trained with reinforcement learning, decides how facts apply to a given request. State-aware reasoning combines both to guide the agent. Experiments on long-term dialogue and memory-use benchmarks show consistent improvements over existing memory systems in retrieval completeness and memory-grounded response accuracy. Code is available.