EviMem: Evidence-Gap-Driven Iterative Retrieval for Long-Term Conversational Memory
For conversational AI systems needing multi-session memory, EviMem provides a targeted iterative retrieval method that significantly boosts accuracy on complex temporal and multi-hop queries.
EviMem improves long-term conversational memory retrieval by explicitly diagnosing evidence gaps, achieving 81.6% Judge Accuracy on temporal questions and 85.2% on multi-hop questions on LoCoMo, outperforming MIRIX with 4.5x lower latency.
Long-term conversational memory requires retrieving evidence scattered across multiple sessions, yet single-pass retrieval fails on temporal and multi-hop questions. Existing iterative methods refine queries via generated content or document-level signals, but none explicitly diagnoses the evidence gap, namely what is missing from the accumulated retrieval set, leaving query refinement untargeted. We present EviMem, combining IRIS (Iterative Retrieval via Insufficiency Signals), a closed-loop framework that detects evidence gaps through sufficiency evaluation, diagnoses what is missing, and drives targeted query refinement, with LaceMem (Layered Architecture for Conversational Evidence Memory), a coarse-to-fine memory hierarchy supporting fine-grained gap diagnosis. On LoCoMo, EviMem improves Judge Accuracy over MIRIX on temporal (73.3% to 81.6%) and multi-hop (65.9% to 85.2%) questions at 4.5x lower latency. Code: https://github.com/AIGeeksGroup/EviMem.