Memory Papers
Memory Architecture

StructMem: Structured Memory for Long-Horizon Behavior in LLMs

AuthorsBuqiang Xu, Yijun Chen, Jizhan Fang et al.

2026
arXivPDF

TL;DR

StructMem uses event-level binding plus cross-event consolidation to build structured long-horizon memory, reaching 76.82 overall on LoCoMo with only 1.937M construction tokens.

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Flat and graph memories break long-horizon reasoning

Existing flat memory systems treat history as an unordered bag of propositions, causing retrieval over long histories to degrade into shallow similarity matching.

Graph memory recovers relational structure but incurs high construction cost, cascading inference, and error accumulation, making long-term temporal reasoning and multi-hop dialogue unreliable.

HOW IT WORKS

StructMem — Event-Level Binding and Cross-Event Consolidization

StructMem centers on Event-Level Binding, Dual-Perspective Extraction, Temporal Anchoring, and Cross-Event Consolidation to build structure-enriched hierarchical memory from conversational streams.

You can think of StructMem like a librarian who first files each event as a timestamped card with both facts and relations, then periodically bundles related cards into higher-level story folders.

This hierarchical organization lets StructMem reconstruct causal chains and temporal dependencies that a plain context window or flat vector store cannot capture, even with extended context.

DIAGRAM

Event-Level Binding and Cross-Event Consolidation Flow

This diagram shows how StructMem processes dialogue turns into dual-perspective event entries, anchors them temporally, and periodically consolidates cross-event structures.

DIAGRAM

LoCoMo Evaluation and Paradigm Comparison

This diagram shows how StructMem is evaluated on LoCoMo alongside flat memory and graph memory paradigms, including ablations without cross-event structure.

PROCESS

How StructMem Handles a Long-Horizon Dialogue Session

  1. 01

    Dual-Perspective Extraction

    StructMem uses Dual-Perspective Extraction with prompts Pfact and Prel to produce factual entries Φi and relational entries Ψi for each utterance.

  2. 02

    Event-Level Binding

    StructMem performs Event-Level Binding by grouping extracted entries into structured episodes that preserve both event content and interactional relations.

  3. 03

    Temporal Anchoring

    StructMem applies Temporal Anchoring by attaching a timestamp τi and embedding ex to each entry, forming temporally grounded event-level units.

  4. 04

    Cross-Event Consolidation

    StructMem periodically runs Cross-Event Consolidation, performing Semantic Event Connections and Memory Consolidation through Synthesis to induce cross-event structure.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Event-Level Binding

    StructMem introduces Event-Level Binding that combines Dual-Perspective Extraction and Temporal Anchoring to preserve factual and relational bindings within each utterance.

  • 02

    Cross-Event Consolidation

    StructMem proposes Cross-Event Consolidation that buffers events, performs Semantic Event Connections via top K retrieval, and synthesizes consolidated memories through LLM prompts.

  • 03

    Paradigm-Level Evaluation on LoCoMo

    StructMem delivers 76.82 overall on LoCoMo with only 1.937M construction tokens, compared to Mem0g’s 68.44 overall and 35.825M construction tokens.

RESULTS

By the Numbers

Overall

76.82

+1.04 over Memobase

Multi

68.77

+0.66 over Zep

Build Tokens (M)

1.937

-33.888 vs Mem0g

Calls

1056

-52458 vs Mem0g

On the LoCoMo benchmark, which tests multi domain, open domain, single session, and temporal reasoning, StructMem achieves the highest overall score while using far fewer construction tokens and API calls. This demonstrates that StructMem’s hierarchical memory can support long horizon reasoning without the heavy overhead of graph memory systems.

BENCHMARK

By the Numbers

On the LoCoMo benchmark, which tests multi domain, open domain, single session, and temporal reasoning, StructMem achieves the highest overall score while using far fewer construction tokens and API calls. This demonstrates that StructMem’s hierarchical memory can support long horizon reasoning without the heavy overhead of graph memory systems.

BENCHMARK

Performance and resource consumption comparison of memory systems on LoCoMo dataset

Overall score on LoCoMo for different memory systems.

KEY INSIGHT

The Counterintuitive Finding

StructMem reaches 76.82 overall on LoCoMo while using only 1.937M construction tokens and 1056 API calls.

This is surprising because graph memory systems like Mem0g consume 35.825M tokens and 53514 calls, yet achieve only 68.44 overall, contradicting the assumption that more structure and tokens always yield better performance.

WHY IT MATTERS

What this unlocks for the field

StructMem shows that hierarchical event centric memory with Cross-Event Consolidation can deliver strong long horizon reasoning without continuous graph maintenance.

Builders can now deploy long term conversational agents that maintain temporal and relational coherence while staying within practical token, latency, and API budgets.

~10 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

Paper: StructMem: Structured Memory for Long-Horizon Behavior in LLMs

Answers use this explainer on Memory Papers.

Checking…