Title: Avoiding LLM Reliance for Freshness Tracking: A Deterministic Framework for Resolving Memory Conflicts

Abstract

As Large Language Model (LLM) based memory systems increasingly store facts that change over time, a persistent challenge arises in conflict resolution: determining which value to return when multiple contradictory entries exist. The FactConsolidation task within MemoryAgentBench (MAB; Hu et al., 2026) highlights this issue by assigning numerical IDs to facts, where counterfactuals possess higher serial numbers, and agents are instructed that newer facts correspond to larger serials. Despite this clear structure, all existing systems demonstrate significant performance gaps. In single-hop scenarios (FC-SH), HippoRAG-v2 achieves only 54%, BM25 scores 48%, Mem0 reaches 18%, and temporal knowledge graphs like Zep/Graphiti manage a mere 7%. Multi-hop tasks remain nearly unsolvable, with performance capping at 7% across 22 different systems.

We identify the root cause as the assembly phase. Baseline approaches typically rely on the LLM to mediate retrieval or generation for conflict resolution, rather than employing version-aware aggregation. Our controlled comparison, utilizing identical backbones, retrieval mechanisms, chunking strategies, and TOP_K settings, demonstrates that substituting the LLM-judgment pipeline with candidate extraction followed by a Python max(serial) operation improves FC-SH scores by 10.8 points for gpt-4o-mini. This advantage expands as data volume increases, growing from an 8-point lead at 6K samples to a 21-point lead at 262K. This improvement represents a holistic pipeline effect, where the resolver, prompt, format, and temperature interact jointly; isolating the resolver’s specific contribution remains a direction for future research.

This deterministic recipe achieves 78.0% on FC-SH using gpt-4o-mini and 94.8% using gpt-4o. For multi-hop tasks (FC-MH), it reaches 30.2% with gpt-4o-mini, rising to 51.5% with gpt-4o through a per-hop deterministic extension of Self-Ask. At the 262K scale, this approach outperforms HippoRAG-v2 by 28 points and surpasses the best published FC-MH result by 20 points. These findings offer a corrective perspective for the field: the primary bottleneck for conflict resolution lies in assembly (post-retrieval aggregation), not storage. Additionally, a LongMemEval knowledge-update test indicates that while the mechanism can be adapted from max(serial) to max(timestamp), it only matches LLM judgment performance (57.8% vs. 64.4%, n=45). Consequently, deterministic aggregation serves as the essential primitive for resolving current-value conflicts and must be integrated with question-type-aware strategies to enhance broader memory QA capabilities.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC