RESEARCH · RESEARCH · #928
EnSIMem: entity-structured long-term memory for agents (arXiv:2609.27279v1)
EnSIMem is a long-term memory architecture that builds dialogue-grounded index entries in the form [entity][entity type][property:value], preserves source turns, timestamps and multimodal fields, and organizes interactions into theme-coherent episodes offline. At runtime it decomposes requests into evidence requirements aligned with the index, retrieves entity-property evidence for point/temporal/compositional/aggregation reasoning, and generates answers from preserved source evidence; the paper reports high answer accuracy and compact, efficient online contexts on long-term agent-memory benchmarks, and code is available on GitHub.
KEY POINTS
- EnSIMem is a long-term memory architecture that builds dialogue-grounded index entries in the form [entity][entity type][property:value], preserves source turns, timestamps and multimodal fields, and organizes interactions into theme-coherent episodes offline.
- At runtime it decomposes requests into evidence requirements aligned with the index, retrieves entity-property evidence for point/temporal/compositional/aggregation reasoning, and generates answers from preserved source evidence; the paper reports high answer accuracy and compact, efficient online contexts on long-term agent-memory benchmarks, and code is available on GitHub.
- Structuring long-term memory by entity-property entries with episode-level provenance enables more precise retrieval and evidence-grounded answers for agents interacting over long time horizons.
WHY IT MATTERS
Structuring long-term memory by entity-property entries with episode-level provenance enables more precise retrieval and evidence-grounded answers for agents interacting over long time horizons.