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

Source: [EviMem: Evidence-Gap-Driven Iterative Retrieval for Long-Term Conversational Memory](<https://www.sota2.com/research/paper/69f433c355a3c32af89fd35f>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-02T05:00:04.163Z

Authors: Yuyang Li, Yime He, Zeyu Zhang, Dong Gong

Published: 2026-04-30T10:37:04.000Z

## 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.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Long-context Conversational Question Answering on LoCoMo Overall](<https://www.sota2.com/research/sota/long-context-conversational-question-answering-on-locomo-overall.md>): Long-context Conversational Question Answering; LoCoMo Overall; G-EVAL: 2.81.

- [Long-context Conversational Question Answering on LoCoMo Single-hop](<https://www.sota2.com/research/sota/long-context-conversational-question-answering-on-locomo-single-hop.md>): Long-context Conversational Question Answering; LoCoMo Single-Hop; G-EVAL Score: 2.98.

- [Long-context Conversational Question Answering on LoCoMo Multi-hop](<https://www.sota2.com/research/sota/long-context-conversational-question-answering-on-locomo-multi-hop.md>): Long-context Conversational Question Answering; LoCoMo Multi-Hop; G-EVAL: 2.89.

- [Long-context Conversational Question Answering on LoCoMo Temporal](<https://www.sota2.com/research/sota/long-context-conversational-question-answering-on-locomo-temporal.md>): Long-context Conversational Question Answering; LoCoMo Temporal; G-EVAL: 3.08.

- [Long-context Conversational Question Answering on LoCoMo Adversarial](<https://www.sota2.com/research/sota/long-context-conversational-question-answering-on-locomo-adversarial.md>): Long-context Conversational Question Answering; LoCoMo Adversarial; G-EVAL Score: 1.94.

- [Long-context Conversational Question Answering on LoCoMo Open-domain](<https://www.sota2.com/research/sota/long-context-conversational-question-answering-on-locomo-open-domain.md>): Long-context Conversational Question Answering; LoCoMo Open-Domain; G-EVAL: 3.17.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.27695>)

[PDF](<https://arxiv.org/pdf/2604.27695.pdf>)
