Source-linked AI summary
Memory Matters More: Event-Centric Memory as a Logic Map for Agent Searching and Reasoning
Yuyang Hu, Jiongnan Liu, Jiejun Tan, Yutao Zhu, Zhicheng Dou
TL;DR
Agent memory systems often use flat storage and shallow semantic retrieval, limiting explicit modeling of logical relations for long-horizon reasoning. CompassMem incrementally builds an event-centric Event Graph and uses it as a logic map for structured navigation. Across LoCoMo and NarrativeQA, it consistently improves retrieval and reasoning performance, especially on multi-hop and temporal tasks.
Problem
Existing agent-memory approaches often use flat representations and semantic matching, while structured methods may omit logical relations needed for long-horizon reasoning.
Method
CompassMem incrementally segments experiences into event units, extracts logical relations, and uses the resulting Event Graph for logic-aware memory search.
Results
CompassMem consistently improves retrieval and reasoning performance across LoCoMo and NarrativeQA, particularly on multi-hop and temporal reasoning tasks.
Takeaways & Limitations
The results support treating agent memory as a structured logic map that actively guides searching and reasoning rather than as flat storage.
Takeaways & Limitations
Event Graph quality depends on the naive LLM-based event-segmentation and relation-extraction pipeline, and evaluation covers only representative benchmarks.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly deployed as intelligent agents that reason, plan, and interact with their environments. To effectively scale to long-horizon scenarios, a key capability for such agents is a memory mechanism that can retain, organize, and retrieve past experiences to support downstream decision-making. However, most existing approaches organize and store memories in a flat manner and rely on simple similarity-based retrieval techniques. Even when structured memory is introduced, existing methods often struggle to explicitly capture the logical relationships among experiences or memory units. Moreover, memory access is largely detached from the constructed structure and still depends on shallow semantic retrieval, preventing agents from reasoning logically over long-horizon dependencies. In this work, we propose CompassMem, an event-centric memory framework inspired by Event Segmentation Theory. CompassMem organizes memory as an Event Graph by incrementally segmenting experiences into events and linking them through explicit logical relations. This graph serves as a logic map, enabling agents to perform structured and goal-directed navigation over memory beyond superficial retrieval, progressively gathering valuable memories to support long-horizon reasoning. Experiments on LoCoMo and NarrativeQA demonstrate that CompassMem consistently improves both retrieval and reasoning performance across multiple backbone models.
1 Introduction
Existing agent memory systems often store experiences as flat text and retrieve them through semantic matching, leaving logical relations and structured navigation underdeveloped. CompassMem addresses this gap with an event-centric Event Graph that guides logic-aware search and reasoning, improving performance on dialogue and long-document benchmarks.
- Most existing methods store memories as independent text segments and rely primarily on simple semantic matching.Some structured approaches still fail to capture relations such as causality and temporal sequences.
- Human memory motivates organizing continuous experience into discrete events connected by rich temporal and semantic associations.Event Segmentation Theory provides the cognitive basis for this organization.
- CompassMem organizes experiences into event units connected by explicit logical relations, including causality and temporal order.Its Event Graph is constructed incrementally through event segmentation, relation extraction, and topic evolution.
- Agents navigate the Event Graph as a logic map to prioritize relevant information, follow meaningful connections, and avoid redundant retrieval.This replaces flat, similarity-based access with structured evidence collection for complex queries.
- CompassMem consistently improves performance on dialogue and long-document benchmarks, particularly for multi-hop and temporal reasoning.The reported evaluation covers LoCoMo and NarrativeQA.
2 Related Work
Prior agent-memory research includes paging, RAG-style, tree-based, graph-based, hierarchical, and compositional organizations. However, memory is still largely treated as passive storage rather than as an explicit logical structure guiding search and reasoning.
- Earlier systems manage long-term memory through paging, segmentation, organization, or lifecycle-management mechanisms.Examples include MemGPT, MemOS, MemoryOS, Mem0, and MemoryBank.
- Recent work introduces tree-based, graph-based, hierarchical, and compositional memory representations to improve organization.These approaches demonstrate benefits from adding structure to memory.
- Despite richer organization, existing methods largely treat memory as passive storage, whereas CompassMem explicitly encodes logical structure and guides searching and reasoning.
3 Preliminary
The preliminary formulation models an agent that incrementally constructs memory from textual observations, retrieves query-dependent subsets, and generates responses conditioned on the query and retrieved memory. The objective is to improve memory construction and retrieval for higher-quality generation.
- At time step t, the agent receives a stream of textual observations, such as dialogue turns or narrative sentences.The observations are represented as text units within X_t.
- The construction process Φ extracts a sub-memory from current observations and integrates it with previously stored memory.This produces the updated memory M(t) from the incoming stream and M(t−1).
- At inference time, the retrieval process Ψ selects a query-dependent subset of the current memory.The selected subset is denoted M(t)|q and is used to answer query q.
- A conditional generation function produces the output y from the query and retrieved memory.
- The formal objective is to design more effective memory construction Φ and retrieval Ψ processes for higher-quality generation.
4 Method
CompassMem incrementally builds a hierarchical Event Graph from coherent events, explicit relations, and evolving topics, then searches it through active, goal-directed navigation. This structure treats memory as a logic map that supports multi-path evidence gathering beyond isolated similarity retrieval.
- Overview: CompassMem organizes experiences as coherent event units connected by explicit logical relations in a hierarchical Event Graph.Events preserve temporal and semantic structure, while relations encode dependencies among them.
- Incremental Hierarchical Memory Construction: The system constructs memory incrementally by segmenting input streams, extracting event relations, and integrating new sub-memories through graph updates.Equivalent events can be merged, related events linked, and otherwise novel events inserted as new nodes.
- Topic Evolution: A topic layer clusters events coarsely to complement fine-grained logical relations and facilitate efficient multi-path exploration.Topics are initialized with clustering and updated online as new events arrive; periodic re-clustering preserves global coherence.
- Active Multi-Path Memory Search: During search, CompassMem retrieves candidate events across distinct topic clusters before selecting starting nodes for graph exploration.This initialization reduces reliance on candidates concentrated in a single semantic aspect of the query.
- Active Multi-Path Memory Search: Explorers navigate event-graph neighborhoods using SKIP, EXPAND, and ANSWER actions while tracking evidence against unsatisfied subgoals.A planner decomposes the query, maintains subgoal progress, and refines the query when collected evidence is insufficient.
- Active Multi-Path Memory Search: CompassMem treats graph topology as carrying logic, so relations constrain exploration paths and guide reasoning over structured dependencies.The search progressively collects distilled evidence rather than retrieving isolated memories solely by similarity.
5 Experiment
CompassMem is evaluated on LoCoMo and NarrativeQA, where it consistently improves reasoning performance across backbones and maintains practical efficiency. Further analyses show that its event-level structure, explicit relations, and search components contribute especially to complex retrieval and reasoning.
- Main Results: CompassMem improves average LoCoMo F1 from 47.92% with HippoRAG to 52.18% on GPT-4o-mini and reaches 52.52% on Qwen2.5-14B.It performs best across LoCoMo subsets, including temporal questions.
- Main Results: CompassMem surpasses CAM by over 5% F1 on GPT-4o-mini and more than 8% F1 on Qwen2.5-14B on NarrativeQA.NarrativeQA requires long-range narrative understanding and evidence aggregation.
- Main Results: Across both benchmarks, CompassMem shows consistent improvements, with the largest gains on tasks requiring complex retrieval and reasoning.The strongest baselines are generally graph-based, while CompassMem adds event-level modeling and logic-aware relations.
- Efficiency Analysis: CompassMem has substantially lower memory construction time than Mem0, A-Mem, and MemoryOS while keeping total processing time and per-question latency practical.Its higher token use is accompanied by substantial performance gains.
- Ablation Study: Removing topic clustering, event units, edges, query refinement, or subgoal generation consistently reduces performance, especially on multi-hop and temporal questions.Single-hop and open-domain questions show smaller degradation.
- Hyperparameter Analysis: Larger topic-selection size p and retrieval size k improve performance, indicating benefits from diverse starting nodes and a broader pool of candidate events.Topic-based selection improves performance over the no-clustering setting, while increasing k yields monotonic gains.
- Model Thinking Ability: CompassMem remains best across task categories with Qwen3-8B, showing that explicit reasoning alone is insufficient without effective memory organization and logic-aware retrieval.Stronger reasoning capacity particularly improves multi-hop and temporal questions across methods.
6 Conclusion
The conclusion presents CompassMem as a structured logic map for agent memory, organizing experiences into connected events to guide searching and reasoning. Results on dialogue and long-document tasks show consistent benefits, particularly for reasoning-intensive tasks.
- Conclusion: CompassMem rethinks agent memory as a structured logic map rather than flat storage.It organizes experiences into coherent events and explicitly models their logical relations.
- Conclusion: Experiments on dialogue and long-document tasks show strong and consistent benefits, particularly for reasoning-intensive tasks.The paper frames this design as supporting memory-guided searching and reasoning.
Limitations
CompassMem’s effectiveness is constrained by the quality of its Event Graph construction and by evaluation on a limited set of representative benchmarks.
- Event Graph quality depends on event segmentation and relation extraction.The paper uses a naive LLM-based pipeline and leaves more fine-grained, robust segmentation for future work.
- The evaluation focuses on a set of representative benchmarks rather than a broader range of tasks and agent settings.Broader evaluation would further strengthen the applicability of CompassMem.
Ethical considerations
The paper studies agent memory architectures using public benchmarks without introducing datasets or intentionally handling identifying personal information.
- The work uses publicly available LoCoMo and NarrativeQA benchmarks and does not introduce new datasets.The authors state that these benchmarks do not contain sensitive personal information.
- The paper states that it does not intentionally collect, infer, or generate content identifying specific individuals.
- Event Segmentation Theory frames human experience as meaningful event units rather than an undifferentiated continuous stream.The theory is used as conceptual background for structuring agent memory.
B.1 Dataset Descriptions
The evaluation uses long-context dialogue and narrative benchmarks, with fixed experimental settings and standardized retrieval procedures for comparable baseline evaluation.
- LoCoMo: LoCoMo evaluates long-range memory and reasoning over very long-term conversations spanning dozens of sessions and hundreds of turns.The experiments exclude adversarial questions and cover single-hop, multi-hop, open-domain, and temporal reasoning categories.
- NarrativeQA: NarrativeQA evaluates reasoning over long narrative documents requiring synthesis across global structure rather than shallow local matching.The evaluation samples 10 documents and 298 QA pairs, with documents averaging around 60,000 tokens.
- Experimental settings: For NarrativeQA, retrieval scope is increased to top-k=10 while other search settings remain unchanged.The adjustment provides broader initial coverage for substantially longer documents.
- Baselines: The compared memory systems include hierarchical, graph-based, interconnected-network, and constructivist structured-memory approaches.Examples include MemoryOS, HippoRAG, A-Mem, and CAM, alongside Mem0’s scalable long-term memory system.
- Baseline implementation: Chunk-based baselines use fixed 512-token chunks and retrieve the top-5 chunks by embedding similarity.Memory-based baselines retain their original settings and implementations to preserve intended behavior and support fair comparison.
C.1 Overall Statistics
CompassMem’s search behavior is generally stable and localized, while reasoning-intensive questions trigger longer exploration and more query refinement.
- Overall statistics: CompassMem processes LoCoMo queries within a moderate, stable time budget without excessive overhead from Event Graph navigation.The median runtime is close to the mean, suggesting consistent behavior across queries.
- Planning and refinement: The Planner generates approximately three subgoals per question, and a high refinement rate indicates iterative query adjustment during search.Not all subgoals are fully satisfied, but partial satisfaction is common given varying evidence availability.
- Item-group analysis: Runtime and exploration depth remain relatively stable across item groups, while refinement and retained evidence vary with reasoning complexity.The reported variation is attributed to differences in item complexity rather than search instability.
- Question categories: Reasoning-intensive multi-hop questions require longer search trajectories, whereas temporal questions show the highest refinement rate despite fewer average steps.Single-hop questions generally require fewer steps and less processing time.
- Exploration paths: Most exploration paths contain two to four steps, while longer paths are rare and associated with more complex queries.This indicates that deep traversal is selectively invoked rather than pervasive.
D Case Study: Multi-hop Reasoning over the Event Graph
CompassMem illustrates logic-aware multi-hop search by decomposing a question into subgoals, navigating connected events, and refining exploration when evidence is incomplete. In the case study, progressively collected events satisfy all subgoals and support an answer derived from distilled evidence.
- Case Setup: The case query requires linking a relocation event with later creative activities mentioned in separate dialogue segments.
- Planner: Subgoal Decomposition: The Planner decomposes the query into three subgoals covering the move, post-move creative activities, and specific artworks.
- Localization: The system retrieves top-5 candidate events, selects starting nodes from distinct topic clusters, and inserts 3 starting nodes into the exploration queue.
- Explorer: Multi-path Navigation and Evidence Collection: Three Explorer agents traverse the Event Graph in parallel, exploring 10 candidate nodes, retaining 7 as evidence, and averaging 2.84-step paths.
- Query Refinement: A single refinement step targets partially supported details, and this mechanism is triggered in 76.4% of LoCoMo questions overall.
- Evidence Aggregation and Answer Generation: The retained events jointly satisfy all subgoals, enabling an answer that paintings and stained glass artworks were created after the move.
- Discussion: The discussion characterizes the process as guided traversal that accumulates evidence across paths, refines gaps, and reasons over event dependencies rather than retrieving one text chunk.