EvoMemState adds RL-trained reasoning state to agent memory
TL;DR
- EvoMemState attaches two states to stored facts: Relational States for fact-to-fact relations, and a Reasoning State for request-time applicability.
- A reinforcement-learning stage converts memory-use experience into a policy that generates the per-request Reasoning State.
- Authors report consistent gains in retrieval completeness and memory-grounded response accuracy on long-term dialogue and memory-use benchmarks, without publishing numbers in the abstract.
EvoMemState treats user memories not as a flat list of facts but as facts plus attached "states" — one for how each fact relates to the others as memory evolves, one for how a fact should be applied to the current request. The framework comes from a new arxiv preprint.
Three stages run end-to-end. Ontology-guided induction pulls out persistent facts and the "Relational States that maintain their validity, scope, and evolving relations." A reinforcement-learning stage "internalizes memory-use experience into a policy through reinforcement learning to generate a request-specific Reasoning State." The final stage "combines the relevant facts with their Relational States and the generated Reasoning State to guide the personal agent."
The authors claim "consistent improvements over existing memory systems in retrieval completeness and memory-grounded response accuracy" on long-term dialogue and memory-use benchmarks. The abstract reports no numbers and names no baselines or specific benchmarks. Code is posted on GitHub.
Originally reported by paper
Read the original article →Original headline: EvoMemState Adds RL-Learned Reasoning State to Personal Agent Memory