MemWeaver: Weaving Hybrid Memories for Traceable Long-Horizon Agentic Reasoning

TL;DR

MemWeaver addresses the challenge of long-horizon reasoning in LLM agents by introducing a consolidation-centric tri-layer memory: a temporally grounded Graph Memory for structured facts, an Experience Memory for reusable patterns, and a Passage Memory for verifiable evidence. It employs a dual-channel retrieval that jointly leverages structured relations and supporting texts to form compact, information-dense reasoning contexts, improving multi-hop and temporal reasoning while drastically reducing input context needs. Empirical results on LoCoMo show MemWeaver outperforming baselines across backbones and tasks, with strong evidence-grounding and robust performance under ablations and varying hyperparameters. The framework advances traceability and cross-session generalization, enabling scalable long-term conversational QA and offering a practical memory architecture for real-world, memory-intensive agents.

Abstract

Large language model-based agents operating in long-horizon interactions require memory systems that support temporal consistency, multi-hop reasoning, and evidence-grounded reuse across sessions. Existing approaches largely rely on unstructured retrieval or coarse abstractions, which often lead to temporal conflicts, brittle reasoning, and limited traceability. We propose MemWeaver, a unified memory framework that consolidates long-term agent experiences into three interconnected components: a temporally grounded graph memory for structured relational reasoning, an experience memory that abstracts recurring interaction patterns from repeated observations, and a passage memory that preserves original textual evidence. MemWeaver employs a dual-channel retrieval strategy that jointly retrieves structured knowledge and supporting evidence to construct compact yet information-dense contexts for reasoning. Experiments on the LoCoMo benchmark demonstrate that MemWeaver substantially improves multi-hop and temporal reasoning accuracy while reducing input context length by over 95\% compared to long-context baselines.

Figures (10)

• Figure 1: Comparison of flat retrieval memory, structured memory, and agentic memory from a reasoning perspective. While flat and structured memories can store and retrieve factual information, they do not explicitly model temporal relations between events, which limits their ability to answer temporally constrained queries. Agentic memory addresses this limitation by maintaining temporally grounded events and reasoning over their order.
• Figure 2: Overview of the MemWeaver framework. MemWeaver integrates Graph Memory (GM), Experience Memory (ExpM), and Passage Memory (PM) into a unified long-term memory system. Dialogue interactions are incrementally written into structured and unstructured memory, while dual-channel retrieval combines relational facts and textual evidence to construct a compact reasoning context for answer generation.
• Figure 3: Hyperparameter sensitivity analysis of MemWeaver. Each heatmap reports performance under different retrieval configurations for (a) Multi-Hop, (b) Temporal, (c) Open-Domain, and (d) Single-Hop questions.
• Figure 4: Human evaluation of evidence support quality for A-Mem and MemWeaver across four question categories. Using GPT-4o-mini as the backbone, human annotators assess, on 25 sampled questions per category, whether the retrieved knowledge provides sufficient support for the generated answers.
• Figure 5: Prompt for entity extraction from a dialogue snippet.