EviMem: Evidence-Gap-Driven Iterative Retrieval for Long-Term Conversational Memory

2026-04-30

AI Summary

This work addresses critical limitations in existing retrieval methods for long-term conversational memory, which struggle with cross-session, time-sensitive, and multi-hop reasoning while lacking explicit diagnosis of missing evidence—leading to blind query refinement. To overcome these challenges, the authors propose EviMem, a novel framework featuring the IRIS closed-loop mechanism that evaluates the sufficiency of retrieved evidence, explicitly identifies information gaps, and drives targeted query reformulation. Complementing this, EviMem introduces LaceMem, a hierarchical memory architecture that organizes conversational evidence from coarse to fine granularity. Evaluated on the LoCoMo dataset, EviMem significantly outperforms MIRIX, improving accuracy on temporal questions from 73.3% to 81.6% and on multi-hop questions from 65.9% to 85.2%, while reducing inference latency by 4.5×.

Abstract

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.

Problem

• long-term conversational memory
• evidence gap
• iterative retrieval
• multi-hop questions
• temporal reasoning

Innovation

• evidence-gap-driven retrieval
• iterative retrieval
• conversational memory
• sufficiency evaluation
• layered memory architecture