EviMem Improves Long-Conversation AI Memory by Diagnosing Missing Evidence

Maintaining coherent memory across extended conversations remains a challenge for AI systems, especially when questions span multiple sessions or require reasoning across disparate facts. Single-pass retrieval often misses relevant context, particularly for temporal queries or those requiring multiple logical steps to answer.

The Problem: Evidence Gaps in Iterative Retrieval#

Current approaches attempt to improve retrieval by refining queries through generated content or document-level signals, but they lack explicit awareness of what information is actually missing. This absence of targeted diagnosis means query refinement remains generic and inefficient, leaving significant evidence gaps unaddressed.

EviMem's Two-Part Solution#

The new system, called EviMem, combines two complementary components:

• IRIS (Iterative Retrieval via Insufficiency Signals): A closed-loop framework that actively detects evidence gaps by evaluating whether current retrieved content sufficiently answers the question, then diagnoses specifically what is missing before driving targeted query refinement.
• LaceMem (Layered Architecture for Conversational Evidence Memory): A hierarchical memory structure that supports progressively refined gap diagnosis, moving from coarse-grained to fine-grained analysis of missing information.

Measurable Performance Gains#

Testing on the LoCoMo benchmark demonstrates substantial improvements over the existing MIRIX baseline:

• Temporal questions: Judge Accuracy improved from 73.3% to 81.6%
• Multi-hop questions: Judge Accuracy improved from 65.9% to 85.2%
• Efficiency: Achieved these gains at 4.5× lower latency

The framework's ability to pinpoint what evidence is missing, rather than blindly refining queries, enables more effective and efficient retrieval during long-term conversations.