PaperlayerView exact revision on arXiv

Source-linked AI summary

BMAM: Brain-inspired Multi-Agent Memory Framework

Yang Li, Jiaxiang Liu, Yusong Wang, Yujie Wu, Mingkun Xu

arXiv:2601.20465v1cs.CL

TL;DR

Long-horizon language-model agents struggle to preserve temporally grounded information and behavioral consistency across sessions, a failure pattern termed soul erosion. BMAM addresses this gap with specialized, complementary memory subsystems, timeline-indexed episodic organization, and fused retrieval signals, achieving 78.45% accuracy on LoCoMo while ablations support the importance of hippocampus-inspired episodic memory for temporal reasoning.

  • Problem

    Long-horizon agents face persistent difficulty preserving temporally grounded information and behavioral consistency across sessions because memory management is limited beyond the current input.

  • Method

    BMAM decomposes agent memory into specialized episodic, semantic, salience-aware, and control-oriented subsystems, with timeline-indexed episodic organization and hybrid retrieval.

  • Results

    78.45% accuracy on LoCoMo, with ablation studies confirming the critical role of hippocampus-inspired episodic memory in temporal reasoning.

  • Takeaways & Limitations

    BMAM provides a general-purpose architecture for persistent memory management and a diagnostic lens connecting long-term memory failures with memory organization choices.

  • Takeaways & Limitations

    Evaluation covers four established long-term memory benchmarks, while broader validation across additional domains remains future work.

Abstract

from arXiv · show

Language-model-based agents operating over extended interaction horizons face persistent challenges in preserving temporally grounded information and maintaining behavioral consistency across sessions, a failure mode we term soul erosion. We present BMAM (Brain-inspired Multi-Agent Memory), a general-purpose memory architecture that models agent memory as a set of functionally specialized subsystems rather than a single unstructured store. Inspired by cognitive memory systems, BMAM decomposes memory into episodic, semantic, salience-aware, and control-oriented components that operate at complementary time scales. To support long-horizon reasoning, BMAM organizes episodic memories along explicit timelines and retrieves evidence by fusing multiple complementary signals. Experiments on the LoCoMo benchmark show that BMAM achieves 78.45 percent accuracy under the standard long-horizon evaluation setting, and ablation analyses confirm that the hippocampus-inspired episodic memory subsystem plays a critical role in temporal reasoning.

1 Introduction

Long-horizon agents struggle to preserve and retrieve temporally grounded, behaviorally consistent information because memory is fragmented and limited by finite context. BMAM addresses this through specialized memory subsystems, timeline organization, hybrid retrieval, and benchmark validation.

  • Long-horizon agents must retain, organize, and retrieve information across diverse tasks, domains, goals, and time scales despite finite context windows.
  • BMAM decomposes memory into episodic, semantic, and salience-aware subsystems coordinated through shared control.
  • Soul erosion describes degradation in temporal coherence, semantic consistency, and identity-related behavior caused by fragmented or misaligned memory.
  • Timeline-indexed organization and hybrid retrieval fuse lexical, semantic, and temporal signals for long-horizon grounding.
  • 78.45% accuracy on LoCoMo demonstrates BMAM’s performance in long-horizon settings, while ablations confirm the hippocampus-inspired subsystem’s critical role in temporal reasoning.
  • Soulfulness is modeled across temporal coherence, semantic consistency, and identity preservation, with benchmark proxies operationalizing these components.

2 Background and Related Work

Prior memory architectures extend context or manage stored information through retrieval, hierarchy, forgetting, or reflection. BMAM is situated among cognitive and agent-centric approaches that emphasize explicit, specialized memory organization for long-horizon behavior.

  • RAG improves factual grounding but leaves retrieved passages implicit and transient rather than reorganizing them into stable internal memory structures.
  • Agent-centric systems introduce explicit memory management through virtual context tiers, forgetting mechanisms, autonomous storage and retrieval, and scalable multi-component stores.
  • Memory-augmented transformers primarily extend token- or sequence-level prediction, whereas BMAM targets what agents store, how they organize it temporally, and how they retrieve it under changing goals.
  • Cognitive-inspired systems separate episodic encoding, semantic consolidation, salience prioritization, event segmentation, and reflective learning across specialized mechanisms.
  • Long-term memory benchmarks provide complementary evaluations of temporal reasoning, preference consistency, and persona recall.

3 BMAM Framework

BMAM coordinates specialized memory subsystems over a unified substrate to manage encoding, consolidation, retrieval, and revision across multiple time scales. Its timeline-indexed episodic organization and hybrid, adaptive retrieval support temporally grounded long-horizon reasoning.

  • Architecture: BMAM uses a central coordinator to route information among specialized subsystems while preserving a unified memory state.The architecture covers storage, retrieval, consolidation, and control functions.
  • Architecture: Episodic, semantic, salience-aware, and control components divide memory responsibilities across complementary functions.Episodic memory stores temporally grounded traces, semantic memory consolidates stable facts and relations, and salience signals influence consolidation and retrieval.
  • Temporal Organization: BMAM organizes episodic memories into timeline-indexed narrative units linked to entities, events, and timestamps, enabling order, duration, and temporal-relation queries.Selected episodic information is consolidated into semantic form to maintain consistency across representations.
  • Memory Lifecycle: BMAM couples fast context access with slower consolidation, procedural stabilization, and iterative coordination across multiple time scales.Its explicit memory loop encodes experiences episodically, maintains a constrained working-memory buffer, and later consolidates selected information into semantic memory and a shared knowledge graph.
  • Retrieval: Retrieval combines episodic, semantic, and timeline evidence using lexical, dense, relational, and temporal signals with weighted reciprocal rank fusion.Temporal evidence supports relative ordering and duration reasoning, while uncertainty and salience can trigger additional retrieval rounds or reweight sources.
  • Memory Revision: Continual learning revises semantic confidence through evidence-based updates while background processes reconsolidate, prioritize, and prune memories.The update rate λ ∈ (0, 1) dampens noisy evidence; pruning reduces uncontrolled growth and obsolete or inconsistent information.

4 Experiments

BMAM is evaluated across complementary long-horizon memory and personalization benchmarks using persistent-memory protocols, baseline comparisons, and component ablations. Results are strongest on LoCoMo and PrefEval, while temporal reasoning, cross-session integration, and exact surface-form matching remain challenging.

  • Experimental Setup: BMAM is evaluated primarily on LoCoMo and LongMemEval, which test conversational memory, temporal reasoning, and consistency over extended interactions.PersonaMem and PrefEval complement these benchmarks by testing persona consistency and preference alignment.
  • Experimental Setup: Persistent memory is reset between independent evaluation units but preserved within each unit, with conversation logs ingested before evaluation and no additional learning during queries.The protocol follows MemOS evaluation metrics and judge prompts while using official baseline scripts.
  • Results: BMAM is strongest on LoCoMo and PrefEval, whereas PersonaMem remains challenging because exact surface-form matching differs from BMAM’s open-ended-generation optimization.PersonaMem also emphasizes shallow persona attributes, while BMAM focuses on temporally grounded identity.
  • Results: 78.45% overall accuracy is achieved on LoCoMo under the MemOS-aligned protocol, with 82.0% single-hop, 70.4% multi-hop, 62.3% temporal, and 79.6% open-domain accuracy.Comparisons with reported baselines are indicative because model backends, prompts, and infrastructure may differ.
  • Results: 67.60% overall accuracy is achieved on LongMemEval, including 100% single-session preference, 87.1% single-session user facts, 59.4% temporal reasoning, and 52.6% multi-session recall.The lower temporal and multi-session scores reflect the difficulty of cross-session temporal integration and explicit time computation.
  • Ablation Analysis: Removing hippocampus-inspired episodic memory causes a 24.62% accuracy drop on a LoCoMo subset, confirming its central role.Removing Prefrontal and Temporal Lobe components instead yields aggregate gains of 5.03% and 4.02% on this single-hop-dominated subset, while Temporal Lobe removal reduces temporal-query accuracy by 12.3%.
  • Discussion: The reported results are means across three runs, and the authors emphasize BMAM’s architectural pattern balancing fast episodic access with control mechanisms.The paper identifies temporal reasoning and cross-session integration as open challenges.
  • Error Analysis: Temporal confusion accounts for 38% of sampled LoCoMo errors, followed by entity ambiguity at 28% and retrieval coverage at 22%.The remaining 12% involve annotation ambiguities or external knowledge, motivating temporal normalization and entity-aware retrieval.

5 Conclusion

BMAM addresses soul erosion by organizing persistent agent memory into specialized subsystems and provides a general-purpose architecture for long-horizon memory management. Experiments show strong LoCoMo performance, while remaining errors motivate further work on temporal, entity, and cross-session memory.

  • 5 Conclusion: BMAM decomposes memory into episodic, semantic, and salience-aware subsystems coordinated through shared control.This modular architecture is presented as a general-purpose approach to persistent memory management and as a diagnostic lens for memory failures.
  • 5 Conclusion: 78.45% accuracy on LoCoMo demonstrates strong performance in long-horizon memory evaluation.Ablation studies also confirm the critical role of hippocampus-inspired episodic memory.
  • 5 Conclusion: Temporal confusion (38%) and entity ambiguity (28%) remain dominant failure modes in BMAM’s error analysis.Cross-session integration is also challenging, reaching 52.6% on multi-session tasks.
  • 5 Conclusion: Future work targets temporal normalization, entity-aware retrieval, and improved cross-session consolidation.The paper also identifies multimodal memory, embodied agents, and adaptive component activation as future directions.

6 Limitations

The evaluation covers four established long-term memory benchmarks, but broader validation across additional domains remains future work.

  • 6 Limitations: The evaluation focuses on four established long-term memory benchmarks, leaving broader validation across additional domains for future work.The benchmarks capture core challenges in long-horizon conversational memory, while the authors separately re-evaluated MemOS under their GPT-4o-mini setting.

7 Ethics Statement

BMAM’s persistent-memory design includes lifecycle, timeline, retrieval, brain-region, and integration components. Responsible deployment requires user control and transparent handling of stored information.

  • 7 Ethics Statement: Persistent memory systems require attention to user consent, data ownership, and long-term data retention.These considerations arise from the handling of information stored across extended interactions.
  • 7 Ethics Statement: Responsible deployment requires transparent memory policies, user control over stored information, and deletion upon request.The paper links these mechanisms to maintaining user trust and complying with applicable privacy requirements.
  • 7 Ethics Statement: BMAM’s memory lifecycle forms a closed loop from perception and encoding through consolidation, reflection, reconsolidation, and forgetting.Forgetting prunes low-salience items, while reconsolidation updates memories when new evidence arrives.
  • 7 Ethics Statement: StoryArc stores per-entity timelines with normalized timestamps to support temporal queries about event order and occurrence.This provides the explicit temporal organization used for questions such as when an event happened or what preceded another event.
  • 7 Ethics Statement: Hybrid retrieval combines lexical, semantic, relational, and temporal pipelines, fusing their results with Reciprocal Rank Fusion.The four components are BM25, dense vectors, a knowledge graph, and StoryArc.

A.3 Extended Ablation Results

Extended ablations identify the hippocampus-inspired subsystem as the most consequential component on the evaluated LoCoMo subset, while showing task-specific and coupled component effects. Additional benchmark analyses reveal strong preference extraction but persistent temporal and multi-session challenges.

  • A.3 Extended Ablation Results: The hippocampus ablation causes the largest accuracy drop, 24.62%, on the 199-question LoCoMo Group 1 subset.This result supports the critical role of episodic memory encoding and retrieval.
  • A.3 Extended Ablation Results: 77.39% full-BMAM accuracy versus 52.76% without the hippocampus, with confidence intervals indicating a statistically significant drop.The reported 95% confidence intervals are 71.1–82.6% and 45.7–59.7%, respectively.
  • A.3 Extended Ablation Results: BMAM achieves 72.9% personalized responses with 0.1% inconsistency violations in PrefEval, while 18.9% preference-unaware violations remain.The results indicate stable preference memory alongside room for improvement in preference detection.
  • A.3 Extended Ablation Results: Only hippocampus ablation differs significantly from Full BMAM in the brain-region study, with p < 0.001.The analysis uses Wilson score confidence intervals and two-proportion z-tests.
  • A.3 Extended Ablation Results: Across the multi-benchmark comparison, BMAM excels on LoCoMo and PrefEval while remaining competitive on LongMemEval and PersonaMem.The radar summarizes the framework’s overall balance across long-horizon dialogue and preference-consistency evaluations.
  • A.3 Extended Ablation Results: BMAM reaches 100% on single-session preference extraction, but temporal reasoning and multi-session integration remain at 59.4% and 52.6%.Within-session recall is 87.1% for SSU and 76.8% for SSA.

A.5 Baseline Comparisons

BMAM is compared with memory-augmented LLM systems across multiple benchmarks, showing strong overall and preference-related performance but weaker temporal results in several evaluations.

  • LoCoMo: 78.45% overall accuracy on LoCoMo outperforms re-run MemOS at 73.90%.BMAM gains 17.5% on single-hop and 13.1% on multi-hop questions, while temporal accuracy is 62.31%.
  • LongMemEval: 100% on LongMemEval single-session preference is the only perfect score, while temporal reasoning and multi-session results lag behind MemOS-1031.Within-session recall remains strong at 76.8% for SSA and 87.1% for SSU.
  • PrefEval: 72.9% personalized rate and 0.1% inconsistency on PrefEval indicate strong preference handling and stable preference memory.The evaluation uses 10 adversarial turns.
  • PersonaMem: 48.9% PersonaMem precision surpasses MemOS at 33.98% and approaches Mem0 at 53.88%.The baseline comparison includes select baselines re-run with official MemOS scripts.

B Case Studies

BMAM case studies show distinct memory mechanisms addressing temporal, semantic, identity, and procedural erosion across long interaction histories.

  • Temporal Erosion: StoryArc temporal indexing resolves chronology when standard retrieval confuses starting a job with leaving it.The case answer places departure from Google around June 2023, after acceptance of the TechStartup Inc. offer.
  • Semantic Erosion: Temporal-lobe consolidation with confidence tracking prioritizes newer dietary statements over outdated preferences.The resulting answer identifies the user as having returned to a fully vegetarian diet.
  • Identity Erosion: Amygdala salience tagging preserves an emotionally significant PhD defense despite many later casual conversations.The retrieved response identifies the thesis as concerning neural memory models and notes its successful defense after five years.
  • Procedural Erosion: Basal-ganglia procedural memory and fixed-point detection preserve consistent coding preferences across sessions.BMAM produces a functional TypeScript React component with Prettier formatting and 2-space indentation.
  • Cross-case synthesis: Different erosion types require different components: StoryArc for temporal, confidence-tracked consolidation for semantic, salience protection for identity, and behavioral patterns for procedural memory.The cases are presented as evidence that no single mechanism suffices.

E Reproducibility Checklist

The reproducibility checklist specifies public benchmarks, evaluation procedures, model and embedding settings, memory-reset rules, and known sources of variation.

  • Artifacts: BMAM uses publicly available LoCoMo, LongMemEval, and PersonaMem datasets, with evaluation scripts following the MemOS protocol.The implementation is stated to be released upon acceptance.
  • Compute and models: Experiments use API-based inference on a single machine without a GPU, with gpt-4o-mini for generation and judging.Embeddings use text-embedding-3-small with 1536 dimensions.
  • Sources of variation: Results may vary slightly because LLM responses are nondeterministic and API model versions may change over time.Embedding and retrieval are deterministic, while generation uses temperature 0.7 and judging uses temperature 0.0.
  • Evaluation state: Memory is cleared between independent test units but preserved within each unit for sequential context.The evaluation therefore retains interaction history within a unit while separating users, groups, or samples.
  • Comparison caveats: Some baseline numbers come from the MemOS paper, select baselines were re-run with official scripts, and minor hyperparameter sensitivity is not fully characterized.These boundaries affect reproducibility and direct comparison.
Loading 2601.20465v1…