EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning

Abstract: LLMs are increasingly deployed as long-term interactive agents, yet their limited context windows make it difficult to sustain coherent behavior over extended interactions. Existing memory systems often store isolated records and retrieve fragments, limiting their ability to consolidate evolving user states and resolve conflicts. We introduce EverMemOS, a self-organizing memory operating system that implements an engram-inspired lifecycle for computational memory. Episodic Trace Formation converts dialogue streams into MemCells that capture episodic traces, atomic facts, and time-bounded Foresight signals. Semantic Consolidation organizes MemCells into thematic MemScenes, distilling stable semantic structures and updating user profiles. Reconstructive Recollection performs MemScene-guided agentic retrieval to compose the necessary and sufficient context for downstream reasoning. Experiments on LoCoMo and LongMemEval show that EverMemOS achieves state-of-the-art performance on memory-augmented reasoning tasks. We further report a profile study on PersonaMem v2 and qualitative case studies illustrating chat-oriented capabilities such as user profiling and Foresight. Code is available at https://github.com/EverMind-AI/EverMemOS.

• The paper introduces a lifecycle-driven memory operating system inspired by engram theory that segments interactions into MemCells and aggregates them into meaningful MemScenes.
• The system demonstrates superior performance on multi-hop and temporal reasoning tasks in benchmarks like LoCoMo and LongMemEval, balancing accuracy with computational efficiency.
• The paper shows that semantic consolidation and dynamic user profiling significantly improve retrieval quality, offering a robust framework for long-horizon LLM agent memory.

EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning

Introduction and Motivation

The increasing deployment of LLMs as persistent interactive agents necessitates robust memory infrastructure for long-horizon reasoning. Traditional approaches, reliant on extended context windows or fragmented retrieval-augmented memory, are fundamentally limited. Context extensions suffer from phenomena such as “Lost-in-the-Middle,” while fragment-based memory systems fail to perform higher-order semantic integration, often resulting in incoherent agent behavior and inconsistent user modeling. EverMemOS presents a system-level paradigm shift, reifying memory as a lifecycle—incorporating ideas from biological engram theory—designed to consolidate fragmented experiences into stable, abstracted representations, and to compose context on a necessity-and-sufficiency basis.

Figure 1: A conceptual comparison between fragment-based and EverMemOS memory in chat interactions.

System Architecture

Engram-Inspired Lifecycle

EverMemOS operationalizes a three-phase memory lifecycle, taking inspiration from biological and cognitive models:

1. Episodic Trace Formation: Continuous interaction streams are segmented into MemCells—atomic units containing narrative episodes, factoid triples, and temporally-scoped Foresight signals (capturing plans/goals with validity intervals).
2. Semantic Consolidation: MemCells are clustered online into MemScenes, forming higher-level semantic structures. Scene-driven aggregation supports user profile evolution, delineating stable from transient states and enabling consistent, longitudinal user modeling.
3. Reconstructive Recollection: Queries trigger agentic, scene-guided retrieval, recomposing necessary and sufficient context using MemScenes and MemCells. Retrieval is augmented by Foresight (time-bounded) and filtered by agentic verification cycles.

   Figure 2: The EverMemOS workflow mapping episodic, semantic, and reconstructive phases, reflecting an engram-inspired memory lifecycle.

Memory Primitives

The MemCell, central to EverMemOS, captures structured, temporally-aware information:

• Episode ($E$): Third-person event narrative.
• Atomic Facts ($\mathcal{F}$): Discrete, factoid statements for high-precision retrieval.
• Foresight ($P$): Prospective intentions or temporary states, annotated by time intervals.
• Metadata ($M$): Timestamps and pointers.

Semantic consolidation is realized via online incremental clustering, updating MemScene representations and enabling dynamic user profiling with recency bias and conflict tracking.

Empirical Evaluation

Reasoning Benchmarks

EverMemOS demonstrates dominant performance on state-of-the-art long-context memory evaluation suites: LoCoMo and LongMemEval. The key empirical findings:

• Superior accuracy on integrated-evidence tasks, specifically multi-hop and temporal reasoning.
• Maintains a favorable accuracy–efficiency trade-off, attaining high performance at moderate token/RAM budgets due to necessity-and-sufficiency retrieval.

  Figure 3: Ablation results indicating the impact of memory structure on overall accuracy in LoCoMo and LongMemEval.

Ablation and Sensitivity Analyses

Ablation studies confirm the architectural value of MemScenes and MemCells:

• Removing MemScenes (flat retrieval) or MemCells (raw dialogue chunks) degrades long-horizon and evidence-integration capabilities.
• Semantic segmentation outperforms heuristic chunking and session boundary oracles, and boosts retrieval quality and answer accuracy regardless of backbone.

Sensitivity analyses on retrieval depth hyperparameters ($N$ for MemScenes, $K$ for episodes) reveal performance saturation at moderate levels, making the system efficient in practice.

Figure 4: Sensitivity of performance to the number of retrieved MemScenes ($N$).

Figure 5: Demonstration of performance versus computational cost as a function of retrieved episode count ($K$).

User Profiling and Qualitative Case Studies

In addition to benchmark QA, EverMemOS advances in user profile consistency and situation-aware foresight, as substantiated by PersonaMem v2. Integrating the consolidated user profile with episodic evidence yields a 9.32-point accuracy gain relative to episodes alone, evidencing nontrivial signal in consolidated abstractions.

Figure 6: Representative case studies demonstrating robust profile extraction, foresight, and episodic context handling in interactive chat.

Theoretical Implications and Future Directions

EverMemOS directly addresses a critical theoretical gap across memory-augmented agent frameworks: the absence of unified, end-to-end lifecycle models for memory construction, abstraction, and reconstructive retrieval. By formalizing memory as a dynamic, multi-phase process, it bridges low-level event capture with semantic distillation and adaptive context composition.

Practically, this enables more interpretable, consistent, and resource-efficient long-horizon agents. The explicit delineation of Foresight, scene-level user modeling, and time-aware filtering extends beyond current state-of-the-art architectures (e.g., Zep, MemOS, Nemori (Nan et al., 5 Aug 2025), Mem0 (Chhikara et al., 28 Apr 2025)). Moreover, the modality-agnostic abstraction of MemCells and MemScenes is compatible with future multi-modal or embodied memory systems, and the modular structure of retrieval and verification cycles is conducive to compositional extension.

Promising avenues for future research include:

• Multimodal extension of MemCells and MemScenes, encompassing audio, vision, and sensor data for holistic agent memory;
• Adaptive, cross-domain profile extraction and management for multi-purpose, personalized assistant agents;
• Open-ended, benchmark-free evaluation regimes focusing on stress-testing memory consolidation and temporal reasoning over ultra-long timelines.

Conclusion

EverMemOS introduces a lifecycle-driven framework for self-organizing memory management in LLM agents, implementing robust engram-inspired primitives for episodic capture, semantic consolidation, and reconstructive context composition. It achieves state-of-the-art accuracy and efficiency on established reasoning and profiling tasks, presents strong architectural justifications for lifecycle-centric memory OS design, and establishes a new direction for structured, long-horizon interactive AI memory systems.