Source-linked AI summary
EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning
Chuanrui Hu, Xingze Gao, Zuyi Zhou, Dannong Xu, Yi Bai, Xintong Li, Hui Zhang, Tong Li, Chong Zhang, Lidong Bing, Yafeng Deng
TL;DR
Long-horizon LLM agents need memory that preserves coherent user models despite limited contexts and fragmented records. EverMemOS introduces an engram-inspired lifecycle that forms MemCells, consolidates MemScenes, and reconstructs query-specific context, achieving state-of-the-art memory-augmented reasoning performance. The evaluation reports strong gains over baselines, while multimodal or embodied use and end-to-end efficiency remain outside the demonstrated scope.
Problem
Existing memory systems often store isolated records, limiting consolidation of evolving user states, conflict resolution, and coherent long-horizon reasoning.
Method
EverMemOS uses Episodic Trace Formation, Semantic Consolidation, and Reconstructive Recollection to transform dialogue into MemCells, organize MemScenes, and compose necessary context.
Results
EverMemOS achieves state-of-the-art performance on memory-augmented reasoning benchmarks, outperforming strongest baselines overall on LoCoMo and LongMemEval.
Takeaways & Limitations
Lifecycle-based organization provides an extensible foundation for more consistent and context-aware interactive agents, with particularly strong gains on multi-hop and temporal questions.
Takeaways & Limitations
Multimodal or embodied extensions are beyond scope, while LLM-mediated memory operations increase latency and computational cost and current benchmarks do not fully stress-test ultra-long timelines.
Abstract
from arXiv · showhide
Large Language Models (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.
1 Introduction
EverMemOS addresses the difficulty of maintaining coherent long-term agent behavior by organizing fragmented experiences into evolving memory structures. Its three-phase lifecycle achieves strong benchmark performance and supports richer chat-oriented behavior.
- Long-term interactive agents must maintain consistent personas and user models while incorporating new constraints across extended interactions.
- Ultra-long contexts can degrade performance through “Lost-in-the-Middle” effects while imposing prohibitive computational costs.
- Existing memory methods often store isolated records without explicitly consolidating fragmented experiences into coherent, stable knowledge structures.
- EverMemOS models memory as a dynamic lifecycle that transforms interaction history into MemCells, consolidates them into MemScenes, and reconstructively retrieves necessary context.
- 9.2% relative improvement on LoCoMo and 6.7% on LongMemEval over the strongest baseline is reported for EverMemOS.
- EverMemOS’s consolidated representation can incorporate new constraints, such as antibiotics use, when producing recommendations, unlike fragment-based memory.
2 Related Work
Prior memory research spans context extension, retrieval and parametric memory, differentiable systems, agent memory, and memory operating systems. EverMemOS distinguishes itself by making lifecycle-based transformation from episodic traces to semantic structures central to long-horizon reasoning.
- Context extension uses sparse attention, recurrence, or length extrapolation, but longer context does not guarantee effective utilization because “Lost-in-the-Middle” persists.
- Retrieval-augmented memory alleviates context-window limits but depends on retrieval quality, while parametric memory can suffer forgetting and instability.
- Early differentiable memory systems introduced external memory interaction but scale poorly and are ill-suited to modern autoregressive LLMs.
- LLM-agent memory systems increasingly integrate persistent episodic, semantic, and hierarchical state, yet many still rely on fragmented text units and limited consolidation.
- Memory operating systems address lifecycle, capacity, fact maintenance, knowledge graphs, or unified scheduling, but their emphases differ from EverMemOS’s reasoning-oriented lifecycle.
3 EverMemOS
EverMemOS implements an engram-inspired lifecycle that converts dialogue into structured memory, consolidates it into thematic scenes, and reconstructs query-specific context. The system also maintains profiles and time-valid foresight for reasoning and chat.
- Framework Overview: The workflow has three phases: Episodic Trace Formation creates MemCells, Semantic Consolidation organizes MemScenes and updates profiles, and Reconstructive Recollection retrieves sufficient context.
- Memory Primitives: A MemCell bundles an Episode, Atomic Facts, Foresight with validity intervals, and Metadata into a temporally grounded memory unit.
- Episodic Trace Formation: Episodic Trace Formation segments continuous dialogue, synthesizes concise third-person narratives, extracts verifiable facts, and generates time-bounded foresight.
- Semantic Consolidation: Semantic Consolidation incrementally assimilates MemCells into existing MemScenes when similarity exceeds threshold τ or creates new scenes otherwise.
- Semantic Consolidation: Scene-level consolidation updates compact user profiles from aggregated summaries, separating stable traits from temporary states through recency-aware updates and conflict tracking.
- Reconstructive Recollection: Reconstructive Recollection selects relevant scenes, filters episodes and expired foresight, verifies context sufficiency, and rewrites queries when supplementation is needed.
- Task Modes: The same retrieval pipeline supports reasoning with Episodes and chat with additional User Profile and time-valid Foresight context.
4 Experiments
Experiments evaluate EverMemOS on long-horizon memory reasoning, profiling, ablations, retrieval budgets, and qualitative conversational capabilities. Results show gains from lifecycle-based organization, semantic consolidation, and efficiency-oriented retrieval.
- Main Results: EverMemOS outperforms the strongest baseline overall on LoCoMo by 7.0% and 9.2%, and on LongMemEval by 6.7%.The strongest baseline is Zep on LoCoMo and MemOS on LongMemEval.
- Main Results: EverMemOS gains 19.7% on LoCoMo multi-hop, 10.0% on temporal tasks, and 20.6% on LongMemEval knowledge updates.The authors associate these gains with MemScenes integrating dispersed evidence and resolving state conflicts.
- Ablation Study: Performance degrades stepwise when MemScenes, MemCells, and external memory are removed from the architecture.Removing MemScenes weakens cross-turn aggregation, removing MemCells forces raw-dialogue matching, and removing external memory collapses long-horizon performance.
- Ablation Study: Semantic episode segmentation outperforms fixed heuristics and Session (Oracle), with accuracy changes of ≤0.7 points across boundary-detection backbones.The comparison is conducted under w/o MemScene to isolate boundary quality.
- Retrieval Scope: Performance gains saturate around N = 10, leading to default retrieval settings of N = 10 MemScenes and K = 10 episodes.These defaults balance performance with computational cost.
- User Profiling: Adding the consolidated User Profile improves PersonaMem-v2 accuracy by 9.32 points over episodes-only, from 43.93 to 53.25.The dataset versions are not directly comparable because their task setups and annotations differ.
- Memory-Augmented Chat: Figure 7 illustrates episode reconstruction, longitudinal profile stability, and experience-grounded foresight in memory-augmented chat.The examples include recalling a Grade-II ankle sprain, tracking waist reduction with stable weight, and using prior failures for planning.
5 Conclusion
The paper concludes that EverMemOS uses an explicit memory lifecycle to support long-horizon LLM agents. It reports state-of-the-art benchmark performance, especially on multi-hop and temporal questions, and presents the system as an extensible foundation for consistent, context-aware interaction.
- Conclusion: EverMemOS models memory through episodic trace formation, semantic consolidation, and reconstructive recollection.The lifecycle is presented as the basis for long-horizon LLM agents.
- Conclusion: EverMemOS achieves state-of-the-art performance on memory-augmented reasoning benchmarks, with particularly strong gains on multi-hop and temporal questions.The conclusion frames the system as an extensible foundation for more consistent and context-aware interactive agents.
Limitations
EverMemOS is evaluated on text-only conversational benchmarks, while multimodal and embodied extensions remain outside the study’s scope. Its LLM-mediated memory operations also increase latency and computational cost, and current benchmarks do not fully isolate ultra-long timeline performance.
- The evaluation covers text-only conversational benchmarks; multimodal and embodied settings are beyond scope.
- LLM-mediated memory construction and retrieval increase latency and computational cost relative to single-pass baselines.Caching, batching, and asynchronous execution may mitigate these costs, but end-to-end efficiency remains future work.
- Current benchmarks lack protocols for stress-testing ultra-long timelines, so the evaluation does not fully isolate performance in those regimes.
A.1 Evaluation Settings and Fair Comparison
The evaluation standardizes answer generation backbones while preserving baseline memory configurations where APIs are available. LongMemEval uses official leaderboard baseline results because extreme input lengths prevent stable end-to-end execution of all baseline APIs.
- Backbone standardization: Final answer generation uses a standardized backbone to isolate memory-management effects from base-model differences.LoCoMo reports GPT-4.1-mini as primary and GPT-4o-mini for comparison with prior work.
- Baseline execution: For Mem0, MemU, MemOS, and Zep, official memory configurations and prompting remain unchanged while the unified backbone is applied only during answering.
- LongMemEval protocol: LongMemEval baseline results come from the official MemOS leaderboard because extreme input lengths prevent stable end-to-end execution of all baseline APIs.EverMemOS is evaluated with GPT-4.1-mini under the same protocol.
- Retrieval configuration: EverMemOS combines dense and BM25 retrieval with Reciprocal Rank Fusion, then applies episode reranking before downstream inference.The default budget retrieves the top-10 MemScenes and selects 10 Episodes.
- Memory construction: MemScenes cluster MemCells, with dataset-specific clustering hyperparameters used across the same pipeline to reflect differing dialogue structures and time spans.A maximum temporal gap limits assignment of a MemCell to a candidate MemScene.
- Evaluation modes: The default quantitative mode uses Episodes-only reasoning; consolidated Profiles and Foresight are evaluated separately.Profiles appear in Table 4, while Foresight is illustrated in the qualitative chat case study.
A.2 LLM-as-Judge Reliability
The reliability study compares LLM-as-judge outputs with human annotations on sampled LoCoMo and LongMemEval questions. Agreement is high, supporting the use of the judging protocol for answer verification, while token accounting documents evaluation costs.
- Human evaluation: 25 Q&A pairs from each benchmark were sampled, and five independent human evaluators judged each generated answer against the question and reference answer.Annotators were recruited through Prolific and compensated at approximately $12.00/hour.
- Agreement: Cohen’s κ exceeds 0.89, with Pearson r of 0.891 on LoCoMo and 0.979 on LongMemEval.The reported results indicate strong agreement between LLM-as-judge and human annotations.
- Interpretation: The study concludes that GPT-4o-mini provides human-level reliability for answer verification, enabling rigorous, reproducible, and cost-efficient evaluation.
- Cost accounting: LoCoMo token usage is logged across 1,540 questions and two backbones, with calls attributed to stages of the EverMemOS pipeline.Profile-related tokens are excluded because the Episodes-only setting does not invoke the Profile module.
A.4 PersonaMem v2: Full Comparison Results
On PersonaMem v2, EverMemOS is compared across 2,447 questions spanning nine scenarios, with profile-like components incorporated when available. It achieves the best overall accuracy, while one-shot retrieval analysis shows that a MemScene budget of 10 already provides strong recall before saturation.
- Full comparison: EverMemOS achieves the best PersonaMem v2 overall accuracy at 53.25%, outperforming MemOS at 50.72% by 2.53 points.The comparison covers 2,447 questions across nine scenarios and uses retrieved memories plus profile-like components when available.
- Retrieval budget: Increasing the MemScene budget initially improves evidence-session recall and answer accuracy, but gains quickly saturate.In the simplified one-shot retrieval setting, N=10 already yields strong recall, motivating avoidance of brute-force scene expansion.
B.2 Accuracy Exceeding Recall on LoCoMo
On LoCoMo, correct answers can exceed annotated-evidence recall because relevant information is redundant, distributed, or inferable from related retrieved content. Increasing retrieval breadth reduces zero-recall cases and narrows the accuracy–recall gap.
- 12–20% of LoCoMo questions are answered correctly even when none of the annotated evidence sessions are retrieved.This is reported as “zero recall” at small K.
- Redundant facts across sessions allow correct answers despite missing annotated evidence, as later sessions restate Caroline’s identity.The annotated session is not necessarily the only session supporting an answer.
- Increasing K from 1 to 3 reduces zero-recall cases by 71% (429→125), narrowing the accuracy–recall gap.
- Distributed activity mentions can support correct answers when retrieved sessions contain sufficient hobby evidence despite missing annotated sessions.The example includes pottery and painting among Melanie’s activities.
- Related signals can support inference-based answers when exact annotated evidence is absent, such as inferring that writing is unlikely from Caroline’s counselor career goal.Retrieved content from other sessions provides the signal used for the inference.
C.2 End-to-End Inference Trace (LoCoMo Multi-Hop Example)
The representative LoCoMo multi-hop trace shows EverMemOS detecting insufficient initial evidence, rewriting queries, and retrieving additional MemCells. The refined retrieval supports an evidence-grounded inference that James likely lives in Connecticut, judged correct by all three judges.
- The trace covers MemBase’s MemScenes and MemCells and a two-round retrieval process with sufficiency checking and query rewriting.
- Round 1 retrieves N=10 MemScenes containing 31 MemCells, but the sufficiency check is false because explicit residence evidence is missing.
- The controller rewrites queries to target James’s residence and location information.
- Round 2 retrieves 40 additional candidates, including evidence that James adopted a dog from a shelter in Stamford.
- The system infers that James likely lives in Connecticut, and the final answer is judged correct by 3/3 LLM judges.