DeepMem: Deep Memory Passing for LLM Agents

ACL ARR 2026 May Submission14774 Authors

Keywords: Agentic Memory，Knowledge Graphs，Memory-Augmented Models

Abstract: Recent agentic memory systems have progressed from flat vector indices to dynamic memory graphs to maintain evolving knowledge during long-horizon interactions, yet they operate under strict shallow topological constraints that limit memory traversals. This rigid single-hop design leaves critical challenges unresolved: it prevents state-changes from propagating to multi-hop dependent notes at write time, and fails to capture deep, mediated relevance at read time. To overcome these limitations, we present \textbf{DeepMem}, a novel framework designed to execute deep memory passing over both the write-time and read-time pathways. Specifically, DeepMem leverages a \emph{Deep Memory Update} strategy to recursively propagate LLM-generated change summaries along graph edges via delta-anchored scoring, and symmetrically utilizes a \emph{Deep Memory Retrieval} mechanism to diffuse query relevance through structural proxies. Comprehensive evaluations on the challenging LoCoMo benchmark demonstrate that DeepMem significantly outperforms state-of-the-art shallow methods, achieving a substantial weighted overall F1 improvement of $+15.7\%$ on Qwen2.5-1.5B-Instruct and $+58.0\%$ on Qwen2.5-3B-Instruct.

Paper Type: Long

Research Area: LLM agents

Research Area Keywords: LLM Agents，Information Retrieval and Text Mining，

Contribution Types: NLP engineering experiment

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: yes

Submission Number: 14774