LightGMEM: Lightweight Agent Graph Memory Generation

ACL ARR 2026 May Submission13678 Authors

Keywords: LLM/AI agents, agent memory, LLM memory

Abstract:

Agent memory systems have been widely proposed to equip large language model (LLM) agents with long-term memory across sessions.
Compared to flat memory, graph-based memory more effectively models relationships between facts, but its construction relies on repeated LLM calls that scale poorly with conversation length.
We propose LightGMEM, a lightweight entity-centric graph memory framework that addresses three key bottlenecks:
(i) replacing per-episode LLM entity extraction with GLiNER2, a zero-shot named entity recognizer, and deferring entity profiling until sufficient cross-episode evidence accumulates;
(ii) introducing conflict-lane partitioning for entity disambiguation, serializing only mentions with overlapping candidate sets while resolving independent ones concurrently;
(iii) adopting Ego-Splitting to construct overlapping communities, allowing entities to participate in multiple retrieval contexts.
On LoCoMo, LightGMEM achieves the best score on 8 of 12 QA metrics with 58.0$\times$ fewer LLM calls and 151.6$\times$ lower construction runtime than Zep. On LongMemEval, it remains competitive on single-session tasks, with trade-offs on update-heavy and temporally reasoning.
These results demonstrate that graph-based memory can remain practical at scale when LLM reasoning is reserved for high-value operations.

Paper Type: Long

Research Area: LLM agents

Research Area Keywords: agent memory

Contribution Types: NLP engineering experiment, Approaches to low-compute settings (efficiency)

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: yes

Submission Number: 13678