Source-linked AI summary
ROAM: Robust Organization of Atomic Memories for Agents through Semantic Relations
Jianjie Zheng, Peng Lai, Sijie Cheng, Jiehui Zhao, Lei Yang, Guanhua Chen
TL;DR
Long-term agents need memory that remains precise as atomic observations accumulate, change, and conflict. ROAM separates semantic relation inference from storage control, organizes memories into Primary and Evidence roles, and fuses details into retrievable Primary views. Across controlled interference and benchmark settings, it consistently outperforms model-based memory-management baselines while preserving historical evidence.
Problem
Atomic observations accumulate redundancy, overlap, and conflicts, while imperative memory mutations risk either redundant storage or loss of historical provenance.
Method
ROAM infers semantic relations, deterministically assigns Primary and Evidence roles, preserves immutable observations, and fuses compatible details into compact Primary read views.
Results
ROAM consistently outperforms model-based memory-management baselines across controlled interference, LongMemEval, and MEME-Post, across three manager models.
Takeaways & Limitations
Separating source retention from retrieval eligibility preserves historical evidence while reducing competition from redundant or superseded records.
Takeaways & Limitations
The evaluation uses benchmark histories and controlled retrieval budgets, so gains should be interpreted within the evaluated settings.
Abstract
from arXiv · showhide
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.
1 Introduction
Long-term agents use atomic memories for precise retrieval, but accumulating observations create redundancy, ambiguity, and conflict. ROAM separates relation inference from deterministic organization, preserves provenance through Primary–Evidence roles, and retrieves fused Primary views; it improves answer quality across evaluation settings.
- Atomic memories support precise retrieval, but streaming observations inevitably create restatements, refinements, and temporal transitions.
- Redundant, overlapping, and outdated records compete with answer-critical evidence, while direct ADD, UPDATE, or DELETE decisions couple semantic interpretation with storage mutation.
- ROAM predicts semantic relations, then deterministically assigns memories to Primary and Evidence roles while preserving supporting or superseded facts.
- Fusion combines compatible details into compact read views, and answer-time retrieval searches only Primary views rather than allowing every atom to compete independently.
- Across controlled competition, LongMemEval, and MEME-Post, ROAM achieves the highest mean accuracy across manager models and consistently outperforms model-based baselines.Relation and fusion ablations show complementary gains, while retrieval diagnostics connect robustness to greater coverage of answer-critical memories.
2 Method
ROAM manages atomic observations through a three-stage extraction, management, and retrieval pipeline. It infers relations against active memories, deterministically preserves immutable records in Primary–Evidence structures, and creates compact fused views for answer-time retrieval.
- Pipeline: The pipeline extracts factual statements, updates structured memory, and retrieves a relevant subset under context budget B.
- Atomic observations: Atomic observations provide fine-grained units for high-precision retrieval, with each record carrying factual text, timestamp, and source context.
- Problem and scope: ROAM addresses redundancy and outdated information by consolidating observations while preserving historical context and provenance.
- Relation inference: For each incoming observation, ROAM infers one of five relations—IND, EQV, OSN, NSO, or CON—against relevant Primary memories.The relations distinguish independence, equivalence, directional entailment, and aligned conflicts.
- Role organization: ROAM stores observations immutably, assigns Primary or Evidence roles, and uses Primary records for later management retrieval while retaining Evidence for provenance.
- Relation-conditioned consolidation: The selected relation determines a deterministic role update, with relation priority protecting state distinctions before consolidating redundant observations.
- Fusion: Fusion changes only the read view: it removes redundancy, preserves compatible specificity, and records temporal changes for answer-time retrieval.
3 Experiments
ROAM is evaluated across controlled interference, natural long-horizon histories, and post-change multi-entity questions. Results show robust gains from relation-guided organization, complementary benefits from fusion, and stronger retrieval coverage under constrained budgets.
- Evaluation settings: The experiments test robustness under retrieval-budget interference, transfer to natural long-horizon and state-update tasks, and the contributions of relations and fusion.The evaluation uses LongMemEval Controlled, full LongMemEval, and MEME-Post.
- Main results: ROAM ranks first across nonzero interference settings and remains consistent across manager models, surpassing the strongest baseline by up to 29.8 percentage points at N = 8.Append-all falls from 87.0% at N = 0 to 24.4% at N = 8.
- Main results: On full LongMemEval, ROAM achieves the highest accuracy under all three manager models, while on MEME-Post it outperforms model-based baselines across managers.LongMemEval accuracies are 63.0%, 59.8%, and 58.6%; MEME-Post reaches 38.9%, 33.1%, and 31.6% across the reported managers.
- Relation ablations: At N = 8, CON-only, CON+EQV, and ROAM achieve 32.1%, 62.3%, and 71.9%, showing increasingly stable performance as more relation types are covered.Contradiction, equivalence, and directional-containment relations each contribute to constructing a reliable active set.
- Fusion ablation: Fusion improves accuracy at every interference level beyond relation-conditioned management alone, indicating complementary benefits from selecting memories and organizing their presentation.The compared variants share identical relation predictions and Primary–Evidence assignments.
- Sensitivity and mechanism analysis: ROAM remains strong across manager sizes and budgets, while retrieval diagnostics show higher coverage of answer-critical sources using fewer average Primary read views.At N = 8, ROAM reaches 69.4%, 71.3%, and 74.5% across manager scales; it represents at least one required source for 95.1% of queries and all required sources for 73.8%.
4 Conclusion
ROAM separates semantic relation inference from deterministic memory control, preserving auditable source histories while exposing compact fused views for retrieval. Across evaluated settings, it consistently outperforms model-based memory-management baselines, with relation-guided active-set control improving answer-critical evidence coverage and fusion adding complementary gains.
- ROAM separates constrained semantic relation inference from deterministic state control while retaining immutable observations in auditable Primary–Evidence structures.
- ROAM exposes compact fused views for retrieval while preserving historical evidence.
- ROAM consistently outperforms model-based memory-management baselines across controlled interference, full LongMemEval, and MEME-Post using three manager models.
- Relation-guided active-set control improves answer-critical evidence coverage and reduces competition from redundant or superseded records, while fusion provides a complementary benefit.
- The results suggest decoupling source retention from retrieval eligibility so historical evidence remains preserved without every observation competing for answer-time context.
Limitations
The evaluation is bounded by benchmark histories and controlled retrieval budgets, and ROAM operates on short extracted factual units rather than richer memory representations. Related work shows that evolving-memory organization has been explored through temporal timelines, knowledge graphs, and semi-structured memory, while the research workflow used reviewed LLM-generated artifacts.
- Benchmark histories and controlled retrieval budgets enable measurable interference and consistent comparisons but cannot capture all deployments.
- Applications may differ in stream length, update frequency, context limits, and conversational complexity, so gains should be interpreted within evaluated settings.
- ROAM operates on short extracted factual units, leaving hierarchical, event-level, and richer structured memories unexplored.
- Related systems address evolving memory through temporal and causal timelines, temporal knowledge graphs, and semi-structured conversational memory.
- LLMs generated controlled confounders, initial relation labels, and answer judgments; all generated labels were manually reviewed and corrected.
C Ethics Statement
This work studies personal factual memory without collecting new personal data, but deployed memory systems may expose sensitive information and retain or propagate incorrect claims. ROAM supports source-history auditability but does not address broader privacy and security requirements.
- The study uses existing benchmarks and involves neither new human interactions nor collection of new personal data.
- Deployed memory systems may expose sensitive information, retain outdated or incorrect claims, and influence later responses.
- ROAM’s Evidence records preserve source history and support auditability but are not a privacy or security mechanism.
- Deployments require consent, access control, secure storage, retention and deletion policies, and mechanisms for users to inspect and correct stored information.
D Evaluation Validation
The evaluation validates both automated answer judgments and relation labels through human review. Human judgments agree with the automated judge on 97.5% of audited examples, while every sampled relation pair is reviewed against ROAM’s definitions.
- 97.5% of 200 audited examples received human judgments agreeing with the DeepSeek V4 Flash answer judge.
- The answer-judge audit spans all three benchmarks and all four compared memory-management methods.
- The relation diagnostic contains 843 stratified memory pairs sampled for approximately balanced coverage of IND, EQV, OSN, NSO, and CON.
- A human annotator reviews every sampled relation pair against ROAM’s relation definitions before diagnostic evaluation.
E Question-Type Results
Tables 6 and 7 report question-type breakdowns for LongMemEval and MEME-Post. Append-all results are pooled across manager-model blocks and repeated in both blocks.
- Tables 6 and 7 report answer accuracy by question type for LongMemEval and MEME-Post.
- Append-all has no manager model, so its two recorded manager-block values are averaged for each question type.
- The pooled Append-all result is repeated in both manager-model blocks, with all values rounded to one decimal place.
F.1 Construction and Confounder Audit
The controlled audit adds fixed semantic and temporal confounders to LongMemEval questions while preserving answer-critical memories. These confounders are designed to test retrieval under plausible competition.
- Construction: The audit retains answer-critical and background memories, then varies the number and type of added confounders across N ∈{0, 2, 4, 6, 8}.The confounder count and type are the only controlled changes across conditions.
- Confounder Types: Type I confounders resemble the query semantically but provide neither the correct answer nor a verifiably incorrect alternative.
- Confounder Types: Type II confounders contain valid but temporally obsolete values and are used only for Knowledge Update questions.Their timestamps precede the corresponding current gold values, representing temporal supersession.
- Audit: The controlled dataset contains 3,760 fixed confounders, including 3,184 Type I memories and 576 Type II memories.These cover 398 non-update questions and 72 Knowledge Update questions.
F.2 Retrieval Diagnostics
ROAM’s retrieval diagnostics measure whether answer-critical sources appear in the final context and how much retained context is attributable to confounders. The analysis uses fixed retrieval outputs and source matching rules.
- Setup: Diagnostics are computed for Gemma 4 12B and Qwen3.5-9B at N = 8 over 470 questions and three runs.Once stores, embeddings, and retrieval orders are fixed, the diagnostic is deterministic.
- Gold-Source Coverage: Gold-source coverage measures whether required gold sources are represented by retrieval units contributing tokens to the final hard-truncated context.For ROAM, the retrieval units are views attached to retrieved Primaries.
- Gold-Source Coverage: Source matching uses exact matches, with a SequenceMatcher ratio of at least 0.65 for Mem0 rewrites; unmapped sources count as uncovered.Metrics are macro-averaged after deduplicating source keys and retrieved matches.
- Token Attribution: The memory block is hard-truncated to 256 Qwen3-8B tokenizer tokens, and confounder-token share counts retained tokens uniquely attributable to fixed confounder sources.Formatting, system-prompt, question, and answer-instruction tokens are excluded.
F.3 Memory-Management Token Consumption
The token-consumption analysis compares management-call overhead under a shared setup. ROAM uses more structured management calls but produces substantially fewer output tokens than Mem0.
- Measurement Setup: Token consumption is measured on LongMemEval Controlled at N = 4 over 470 episodes using the same manager model and counting convention.Measurements include relation inference, memory-operation decisions, and fusion, but exclude extraction, answering, and evaluation.
- ROAM Overhead: ROAM decomposes management into relation inference, deterministic state updates, and optional fusion, resulting in more management calls.
- ROAM Overhead: 1.90k output tokens per episode is 83.4% fewer than Mem0, consistent with ROAM’s constrained relation-label outputs.
- ROAM Overhead: Shared prefixes account for 94.6% of ROAM’s input tokens before reuse.The analysis identifies repeated instructions as a substantial portion of input volume.