Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs

ICLR 2026 Conference Submission20312 Authors

Keywords: long-term memory, conversation, retrieval-augmented generation

TL;DR: We introduce BEAM, a multi-domain benchmark of long (100K–10M token) conversations with comprehensive memory probes, and LIGHT, a cognitive framework that improves LLM memory by 3.5%–12.69% over the strongest baselines.

Abstract:

Evaluating the abilities of large language models (LLMs) for tasks that require long-term memory and thus long-context reasoning, for example in conversational settings, is hampered by the existing benchmarks, which often lack narrative coherence, cover narrow domains, and only test simple recall-oriented tasks. This paper introduces a comprehensive solution to these challenges. First, we present a
novel framework for automatically generating long (up to 10M tokens), coherent, and topically diverse conversations, accompanied by probing questions targeting a wide range of memory abilities. From this, we construct BEAM, a new benchmark comprising 100 conversations and 2,000 validated questions. Second, to enhance model performance, we propose LIGHT–a framework inspired by human cognition that equips LLMs with three complementary memory systems: a long-term episodic memory, a short-term working memory, and a scratchpad for accumulating salient facts. Our experiments on BEAM reveal that even LLMs with 1M token context windows (with and without retrieval-augmentation) struggle as dialogues lengthen. In contrast, LIGHT consistently improves performance
across various models, achieving an average improvement of 3.5%–12.69% over the strongest baselines, depending on the backbone LLM. An ablation study further confirms the contribution of each memory component.

Primary Area: datasets and benchmarks

Submission Number: 20312