Remembering Beyond the Context Window: The Capacity, Consistency, and Control Gaps in Long-Term Memory for LLM Agents
Year: 2026
Type: article
Abstract: Large language model (LLM) agents are increasingly expected to remember users, tasks, and outcomes across sessions, and a dedicated memory infrastructure has emerged to meet that expectation. This survey reviews 21 core sources (2020–2025) on agent memory through an original three-gap framework—capacity (how much is retained), consistency (whether what is retained remains correct), and control (whether retention can be governed)—drawing only on peer-reviewed or publicly verifiable sources for every reported figure. The architecture side has matured rapidly across three generations: operating-system-inspired prototypes, production memory pipelines, and neurobiologically inspired retrieval. The evidence side has not kept pace. Effective context length measured by controlled benchmarks falls far below nominal context length, positional effects corrupt mid-context recall, external memories routinely conflict with parametric knowledge, and knowledge-editing procedures fail to propagate beyond the edited fact. Unlearning remains fragile exactly where retention requirements make it legally consequential. We argue that memory must be treated not as a performance feature but as an auditable record, and map the three gaps onto evaluation standards, privacy regulation, and the emerging requirements for inspectable memory lifecycles.
Authors: Deng, Weibei, Song, Xinyu
Institution: Geely University of China
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7220027987
