PaperlayerView exact revision on arXiv

Source-linked AI summary

APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI

Pratyay Banerjee, Masud Moshtaghi, Shivashankar Subramanian, Amita Misra, Ankit Chadha

arXiv:2604.14362v1cs.CLcs.AIcs.IR

TL;DR

Long-term conversational memory remains vulnerable to noise and unreliable retrieval over evolving information. APEX-MEM uses a temporally grounded property graph with append-only event storage and a multi-tool retrieval agent, achieving strong results across conversational-memory benchmarks, including 88.88% on LOCOMO and 86.2% on LongMemEval.

  • Problem

    Long-term conversational memory systems struggle with noise from extended contexts and unreliable answer accuracy from naïve retrieval, especially when information evolves or conflicts.

  • Method

    APEX-MEM structures conversations as temporally grounded events in a property graph, preserves information append-only, and resolves relevant facts at query time through entity linking, GraphSQL, and hybrid search.

  • Results

    APEX-MEM achieves state-of-the-art performance across multiple benchmarks, including 88.88% accuracy on LOCOMO and 86.2% on LongMemEval.

  • Takeaways & Limitations

    The results support property graphs as a foundation for preserving rich conversational information while enabling retrieval and conflict resolution across varied memory tasks.

  • Takeaways & Limitations

    Graph construction is computationally costly, and performance depends on accurate entity and property resolution and on the QnA agent’s tool-use capabilities.

Abstract

from arXiv · show

Large language models still struggle with reliable long-term conversational memory: simply enlarging context windows or applying naive retrieval often introduces noise and destabilizes responses. We present APEX-MEM, a conversational memory system that combines three key innovations: (1) a property graph which uses domain-agnostic ontology to structure conversations as temporally grounded events in an entity-centric framework, (2) append-only storage that preserves the full temporal evolution of information, and (3) a multi-tool retrieval agent that understands and resolves conflicting or evolving information at query time, producing a compact and contextually relevant memory summary. This retrieval-time resolution preserves the full interaction history while suppressing irrelevant details. APEX-MEM achieves 88.88% accuracy on LOCOMO's Question Answering task and 86.2% on LongMemEval, outperforming state-of-the-art session-aware approaches and demonstrating that structured property graphs enable more temporally coherent long-term conversational reasoning.

1 Introduction

APEX-MEM addresses unreliable long-term conversational memory by structuring evolving information as temporally grounded events and resolving it at retrieval time. Its multi-tool graph retrieval combines entity linking, structured traversal, and hybrid search to reduce noise and support temporal reasoning.

  • Longer context windows can trade memory for noise, increasing irrelevant facts and hallucinations, while retrieval accuracy does not reliably ensure answer accuracy.
  • Structured memory representations organize conversational knowledge to improve retrieval precision, preserve coherence, and enable relational and temporal reasoning.
  • Existing entity-centric and consolidation-based systems may miss nuanced attributes, temporal evolution, and contextual details needed for conflict resolution.
  • APEX-MEM combines a domain-agnostic property graph, append-only event storage, and retrieval-time temporal resolution for conversational memory.Facts remain anchored to events so contradictions and revisions can be resolved when queried rather than overwritten during construction.
  • APEX-MEM retrieval combines canonical entity lookup, GraphSQL traversal, and hybrid search to access relevant graph context and support structured temporal reasoning.GraphSQL supports validity tracking, temporal ordering, duration calculations, and multi-hop traversal across events.

2 Related Work

Conversational memory research progressed from larger contexts and basic retrieval toward explicit memory architectures, including hierarchical, agentic, and entity-centric systems. These approaches improve selected capabilities but expose trade-offs in factual retention, temporal coherence, complexity, and noise sensitivity.

  • Early Approaches to Conversational Memory: Earlier context-window and RAG approaches suffered from noise, with adversarial performance dropping to 15.7% F1 despite larger context and RAG offering modest gains.
  • First-Generation Memory Systems: Explicit memory systems addressed these weaknesses through memory hierarchies, paginated summaries, or psychologically motivated updates, but reported limited conversational F1 results.MemGPT, ReadAgent, and MemoryBank reported 26.65%, 9.15%, and 5.0% F1, respectively, in the cited comparisons.
  • Advanced Memory Architectures: Recent architectures use autonomous linking, hierarchical retrieval, or entity-centric graphs to improve multi-hop and single-hop performance while introducing different complexity and latency trade-offs.
  • Figure 1 depicts APEX-MEM as an end-to-end pipeline from unstructured conversation to a GraphSQL question-answering agent.
  • Advanced Memory Architectures: MIRIX reached 85.4% accuracy with six specialized memory stores and multi-agent routing, but its eager-update strategy risks losing nuanced temporal information.

3 APEX-MEM Graph Construction

APEX-MEM constructs a directed property graph by incrementally merging document-derived subgraphs under a temporal event ontology. Schema-constrained extraction and retrieval-augmented resolution preserve typed, time-bounded facts with provenance for later reasoning.

  • APEX-MEM represents nodes and typed edges with ontology-assigned labels and key-value properties in a directed property graph.Nodes can include entities, events, and facts, while properties encode structured information.
  • The system incrementally merges document-derived subgraphs using soft canonicalization, fusing candidate entities and properties when resolution criteria are met.
  • 3.1 Ontology: Its ontology defines 35 entity classes and represents facts as confidence-scored assertions with temporal validity intervals, evidence, and links to conversational events.Events include timestamps, locations, participants, associated facts, and supporting textual evidence.
  • 3.2 Entity and Property Resolution: Entity and property resolution retrieves top-k dense-search candidates, then uses structured LLM reasoning to choose an existing item, propose a new one, or return none.The pipeline also normalizes names and property schemas while retaining confidence scores and rationales.
  • 3.3 Fact Extraction: Few-shot, schema-constrained extraction converts conversational turns into typed event representations with normalized timestamps, confidence scores, and evidence linked to source utterances.

4 Graph Agents

APEX-MEM uses a ReAct-style graph agent that combines schema inspection, entity retrieval, SQL-based graph reasoning, and hybrid search to answer temporally grounded questions. Tool outputs are accumulated across steps, enabling temporal resolution and multi-stage reasoning over the graph.

  • Graph Agents: The ReAct-style agent selects schema inspection, entity lookup, GraphSQL, Search, or answer actions while accumulating tool outputs across reasoning steps.Temporal references are resolved into dates and date ranges before tool invocation.
  • SCHEMAVIEWER: SCHEMAVIEWER exposes the graph schema and usage guidance, serving as a meta-level planning aid for query and temporal-reasoning strategies.It can return examples and instructions alongside the relational schema.
  • ENTITYLOOKUP: ENTITYLOOKUP maps free-text queries to ranked entity documents using hybrid retrieval, then returns time-aware facts and property snapshots grounded in event dates.The tool canonicalizes surface forms to graph identifiers and exposes temporal anchors for downstream reasoning.
  • GRAPHSQL: GRAPHSQL provides validated read-only SQLite queries over graph tables, supporting precise joins, aggregations, mathematical operations, and temporal computations.The interface permits one safe SELECT or WITH ... SELECT statement and forbids updates and DDL.
  • SEARCH: SEARCH maps a question to candidate entities, properties, events or evidence, and conversation turns, producing a relevant subgraph for subsequent filtering or aggregation.It combines graph, relational, semantic, and entity-property retrieval capabilities.

5 APEX-MEM Online Construction

For very long conversations, APEX-MEM constructs the graph online rather than materializing all documents offline. It first identifies question-relevant documents and builds the graph only from their temporally ordered subset.

  • 5 APEX-MEM Online Construction: APEX-MEM limits online graph construction to temporally ordered documents whose relevance to the input question exceeds a threshold.Relevance is determined using semantic and lexical search over the document set.
  • 5 APEX-MEM Online Construction: Online construction avoids building a complete graph when very long conversations contain many documents irrelevant to expected user questions.The method is motivated by settings where the document count is more than 10^3.

6 Experiments and Analysis

APEX-MEM is evaluated across conversational-memory and difficult search-augmented QA benchmarks, with ablations testing tool contributions and results showing strong performance and broad generalization.

  • 6.2 Experimental Setup: The experiments use LLM-as-a-Judge evaluation across benchmarks, with construction and tool configurations specified for different datasets.Judge scores are averaged over three trials, with reported variation below ±1 standard deviation.
  • 6.3 Results: 88.88% overall accuracy on LOCOMO and 86.2% on LongMemEval surpass prior best systems and session-aware RAG baselines.On LOCOMO, APEX-MEM exceeds MIRIX by 3.50 points; on LongMemEval, it exceeds Nemori by 11.6 points and session-aware RAG by 13.7 points.
  • 6.3 Results: Adding Search after SchemaViewer, EntityLookUp, and GraphSQL raises LOCOMO accuracy from 79.45% to 87%.GraphSQL improves the two-tool system from 77.19% to 79.45%, while Search adds a further 7.55-point gain.
  • 6.3 Results: GraphSQL-only retrieval requires 27,282 tool calls versus 8,260 for the hybrid approach, which reaches higher accuracy with more balanced tool usage.The paper reports GraphSQL-only systems using 3.3x more tool calls while achieving 79.45% accuracy, compared with 87% for the hybrid approach.
  • 6.3 Results: 40.15% accuracy on SealQA-Hard exceeds O3 at 34.6% and other research agents under conflicting and noisy web-search conditions.The strongest-baseline improvement is 5.55 percentage points.
  • 6.3 Results: APEX-MEM maintains less than 5 percentage points of variation across LOCOMO question types and performs strongly across LOCOMO, LongMemEval, and SealQA-Hard.The evaluation covers long-term conversational memory, extremely long inputs, and conflicting or noisy web-search evidence.

7 Conclusion

APEX-MEM combines ontology-supported property graphs, append-only temporal storage, and retrieval-time resolution to improve long-term conversational memory. It achieves strong results across benchmarks while leaving challenges in noisy multi-document reasoning and broader task coverage.

  • Conclusion: APEX-MEM achieves 88.88% on LOCOMO Question Answering and 86.2% on LongMemEval, surpassing prior systems and session-aware RAG baselines.It exceeds MIRIX by 3.50 points on LOCOMO and improves over session-aware RAG baselines by 13.7 points on LongMemEval.
  • Conclusion: APEX-MEM combines append-only construction with retrieval-time resolution using entity linking, GraphSQL traversal, and hybrid search.This design preserves temporal information while supporting conflict resolution and efficient retrieval.
  • Conclusion: The system is evaluated on conversational question answering across LOCOMO-QA, LongMemEval, and SealQA-Hard rather than generative narrative or multimodal tasks.Event summarization and multimodal dialog generation are identified as future work.
  • Conclusion: SealQA-Hard remains challenging, with 40.15% accuracy on noisy, conflicting multi-document scenarios.The authors identify more complete fact extraction and better handling of implicit relationships, temporal nuances, and contextual dependencies as needed improvements.

Limitations

The main limitations concern construction cost, ontology coverage, dependence on capable tool-using base models, and efficiency on difficult or multimodal settings.

  • Limitations: Graph construction is computationally expensive because entity resolution and property extraction rely on large language models.Resolution errors or ambiguities can propagate through the graph and affect downstream retrieval and reasoning.
  • Limitations: The domain-agnostic ontology may miss domain-specific nuances, so specialized queries may require extended or customized schemas.The current ontology contains 35 entity classes.
  • Limitations: APEX-MEM depends on the QnA agent’s ability to generate correct SQLite queries and use tools effectively.With GPT4o, LOCOMO accuracy was 86.35% versus 88.88% with GPT5, despite prompt additions for query errors.
  • Limitations: The graph agent requires multiple tool invocations, which can increase response latency, while performance remains limited on noisy conflicting documents.SealQA-Hard accuracy is 40.15%.
  • Limitations: The current implementation is limited to text-based interactions, leaving multimodal inputs such as images, audio, and video for future work.The authors also identify hybrid integration with long-context and parametric-memory systems as future directions.

Ethical Considerations

The paper reports licensing review, de-identification, privacy safeguards, and limited transparent use of generative AI during development and evaluation.

  • Ethical Considerations: The authors reviewed datasets and computational models to confirm that their licenses permit academic research and publication.This review covered the scientific artifacts used in the research.
  • Ethical Considerations: All experimental datasets were de-identified to protect participant anonymity and privacy.The stated purpose was maintaining anonymity and protecting individual privacy.
  • Ethical Considerations: Generative AI was used in limited, transparent roles for language refinement, code test cases, and automated judging that supplemented human evaluation.The authors distinguish these uses from the core research contributions.

A.1.1 Tool Usage Patterns and Strategic Differences

APEX-MEM’s hybrid retrieval strategy uses complementary tools to improve accuracy and efficiency, while GraphSQL alone is especially strong for temporal reasoning but requires more calls.

  • Tool Usage Patterns and Strategic Differences: APEX-MEM uses 8,260 GraphSQL, 8,900 Search, and 5,160 EntityLookup calls, whereas GraphSQL-only invokes GraphSQL 27,282 times.The ablation therefore uses 3.3× more GraphSQL calls than the full system.
  • Tool Usage Patterns and Strategic Differences: 87.00% LOCOMO accuracy for APEX-MEM exceeds the GraphSQL-only ablation’s 79.45% by 7.55 points.The advantage reaches 11.18 points on open-domain questions, where semantic Search retrieves information beyond structured queries.
  • Tool Usage Patterns and Strategic Differences: GraphSQL-only performs competitively on temporal reasoning, reaching 82.29% versus APEX-MEM’s 79.17%.SQL temporal operators and date functions provide natural expressiveness for these queries.
  • Tool Usage Patterns and Strategic Differences: The hybrid architecture balances structured queries, entity retrieval, and semantic search to handle diverse reasoning requirements more efficiently.Its strategic tool selection reduces the need to express every reasoning operation through SQL queries.
  • Tool Usage Patterns and Strategic Differences: APEX-MEM reaches approximately 84–86% accuracy with 10 tool calls across most categories, while answering 80–90% of questions within 10–20 calls with Claude 4.5 Sonnet.GraphSQL-only improves more gradually and generally requires about 20–30 calls to become competitive.

B.2 Adversarial Questions: The Noise Introduction Effect

On adversarial questions, additional tool calls can introduce noise rather than improve accuracy, with most agents reaching a performance ceiling near 20 calls. The full hybrid APEX-MEM architecture remains stronger and more efficient than single-tool variants on broader reasoning tasks.

  • Adversarial Questions: At 40 tool calls, agents converge to approximately 80–85% accuracy on adversarial questions after performance stabilizes beyond 20 calls.At 20 calls, Sonnet reaches 85%, Haiku 80%, and GraphSQL 78%.
  • Tool Ablations: The full APEX-MEM system reaches 87–91% accuracy across most categories, whereas GraphSQL-only plateaus at 79–83% and struggles on open-domain questions.GraphSQL remains competitive on temporal queries, reaching 82% accuracy, but lacks Search for broader contextual retrieval.
  • Tool Ablations: APEX-MEM achieves higher accuracy with fewer calls because EntityLookup, GraphSQL, and Search combine direct lookup, structured reasoning, and semantic retrieval.GraphSQL-only requires 20–30 calls to reach levels APEX-MEM achieves with 10–15 calls.
  • Execution Robustness: GraphSQL execution succeeds most often with Claude Sonnet 4.5 at 97.6%, while the multi-tool architecture recovers 87% of SQL failures through schema review and fallback tools.Recovery uses SchemaViewer, EntityLookup, and Search when SQL execution fails or graph structure is insufficient.

I Ontological Architecture

APEX-MEM organizes diverse information into a layered, ontology-grounded property graph that preserves temporal order, events, facts, entities, and semantic types. This structure supports canonicalized cross-domain knowledge extraction, temporal reasoning, and unified querying.

  • Architecture Overview: APEX-MEM transforms diverse information sources into a typed knowledge graph through five layers: temporal, event, knowledge, entity, and semantic.The architecture represents both structural relationships among data components and semantic type instantiation.
  • Temporal Layer: The temporal layer represents sessions and turns with timestamps, preserving chronological order across conversations, news streams, and document events.This ordering supports interpretation of information context and evolution over time.
  • Event and Knowledge Layers: The event and knowledge layers extract temporally anchored events and convert them into subject-property-value facts with validity intervals and confidence scores.Evidence spans maintain traceability for extracted events, while typed properties support consistent querying.
  • Entity Layer: The entity layer canonicalizes named entities across contexts, assigning unique identifiers, normalized names, and semantic classifications.Entity resolution links mentions of the same entity across conversations, news, and documents.
  • Semantic Layer: The semantic layer provides shared types such as PERSON, CREATIVE_WORK, PLACE, PRODUCT, GROUP, TOPIC, EVENT, and NATURAL_PHENOMENON for cross-domain reasoning.The graph structure connects conversational turns and events to entities through participation relationships.
Loading 2604.14362v1…