Agents that interact with users over extended periods need to recall facts, preferences, events, and changes from a continuously growing dialogue history. Existing memory systems usually compress interactions into generic summaries or return anonymous text chunks, making it hard for the agent to pinpoint the correct entity, property, and supporting evidence. EnSIMem introduces an entity‑structured long‑term memory architecture that addresses these issues.
During offline construction, the system first partitions the dialogue into theme‑coherent episodes, then creates dialogue‑grounded index entries of the form [entity][entity type][property:value]. Each entry retains the original turns, timestamps, and any multimodal fields (e.g., images, audio), providing fine‑grained provenance.
In online interaction, the agent's request is decomposed into evidence requirements that align with the indexed properties. An entity‑property lookup followed by adaptive retrieval gathers the evidence needed for point, temporal, compositional, and aggregation reasoning. The agent then generates its response directly from the preserved source evidence instead of lossy memory summaries.
On long‑term memory benchmarks, EnSIMem achieves high answer accuracy while keeping contexts compact and online latency low. The results demonstrate that entity‑structured indexing and episode‑level provenance offer a reliable foundation for long‑term memory in agents. The implementation is open‑source at https://github.com/RamonMeng/EnSIMem.
Review: EnSIMem’s transformation of dialogues into queryable entity‑property tables enables efficient and interpretable long‑term retrieval, presenting a practical blueprint for building conversational agents with persistent memory capabilities.