ROAM: Robust Organization of Atomic Memories for Agents through Semantic Relations

2609.09778 ↗ · 2026-09-09 · cs.CL

Jianjie Zheng, Peng Lai, Sijie Cheng, Jiehui Zhao, Lei Yang, Guanhua Chen

21 pages, 4 figures, 10 tables

The finding

Headline claim

Replacing direct add/update/delete memory operations with explicit relation classification plus deterministic role assignment and fusion improves long-term agent QA accuracy, with gains growing as interfering memories accumulate.

The test

Three evaluation settings: LongMemEval Controlled (470 questions, 0 to 8 added confounders), full 500-question LongMemEval, and 694 MEME-Post questions, with three manager models (Qwen3.5-9B, Qwen3-8B, Gemma-4-12B) under a 256-token memory budget, against Append-all, Mem0, and EverMemOS-style baselines.

Effect size & scope

Highest mean accuracy at every nonzero interference level and on both natural benchmarks; up to 29.8 percentage points over baselines (e.g., ROAM 76.2 vs Append-all 25.4 at N=6 with Gemma-4-12B). Mechanism analyses report 15.6 points higher answer-critical source recall and 11.5 points lower confounder-token share. At N=0 the advantage nearly vanishes (ROAM 84.7 to 87.2 vs Append-all 87.0), so the gains are specific to redundancy and conflict handling, not base answering.

Limitations

The authors confine their claim to memory management under a fixed 256-token budget and same extraction/retrieval pipeline; gains depend on relation-inference quality from small open-weight managers, and their controlled benchmark is self-constructed from LongMemEval rather than independent. Unstated: the EverMemOS-style baseline is a partial reimplementation excluding trace and profile machinery, answer judging relies on DeepSeek-V4-Flash, and no error bars or significance tests are shown in the provided text.

Counterpoint

A controlled confounder-count setting designed by the authors can favor a method explicitly built to suppress confounders, and the up-to-29.8-point headline is taken at the most extreme interference level where Append-all catastrophically fails; on the least-constructed benchmark (N=0) differences are within a point or two, so the result may largely restate that unmanaged stores drown retrieval under heavy redundancy, which is already well known.

Economy implication

Little or nothing yet in direct economic terms. Modestly, if agent memory reliability keeps improving at near-zero marginal algorithmic cost, it supports the reliability leg of agent ROI (B1) but this paper measures benchmark QA accuracy, not enterprise task completion or cost.

Who this matters for

Technical Capability: Agent builders get concrete evidence that decoupling semantic relation inference from storage mutation beats imperative memory management, with ablations identifying which relations matter.

Abstract

Long-term language-model agents rely on external memory across interactions. Atomic memories are particularly useful: their fine-grained semantic boundaries enable precise retrieval and direct comparison between observations. Yet accumulating atoms inevitably become redundant, overlapping, or conflicting. Existing methods often ask an LLM manager to add, update, delete, or rewrite memories directly, coupling semantic interpretation, storage decisions, and content generation in one error-prone operation. We introduce ROAM, a relation-guided framework that uses atomicity for management while allowing richer answer-time representations. ROAM classifies incoming--stored atom pairs as independent, equivalent, directionally subsuming, or conflicting, then organizes observations into active Primary and supporting Evidence roles. Fusion subsequently combines complementary details and temporal changes into compact, potentially non-atomic views. Only Primary views are retrieved for answering, preventing redundant or outdated atoms from competing independently. Across models and evaluation settings, ROAM improves answer accuracy by up to 29.8 percentage points. Ablations show complementary benefits from different relations and consistent gains from fusion beyond role organization. Mechanism analysis further finds 15.6-point higher answer-critical source recall and an 11.5-point lower confounder-token share. ROAM remains robust across manager scales.