Controlled Memory Interference for Continual LLM Agents

Long-term memory is crucial for AI agents to maintain continuity, personalize behavior, and learn from accumulated experience across sessions. However, memory evolution is not merely about storing more information; new experiences can reinforce, revise, or interfere with existing memories. Current systems primarily focus on memory construction and relevance-based retrieval, often overlooking the complexities of multiple relevant memories differing in state, temporal validity, or authority. This research introduces Controlled Memory Interference (CMI), a diagnostic and data-generation framework designed to systematically study how agent memory evolves under different memory relationships. Through controlled memory evolution experiments, it was found that while benign accumulation has limited effects, relationship-specific interference sharply suppresses an agent's ability to update its memory, with minimal gains in stability. This interference can occur either by blocking exposure to the target memory or by disrupting its subsequent use. The study also revealed distinct interference pathways for lexical and dense retrieval methods, and that "poisoning" attacks are more sensitive to update-authority cues than to recency. Beyond diagnosis, CMI provides targeted examples for developing interference-aware memory learning techniques, which can improve the distinction between valid updates and interference-inducing memories while preserving performance on original tasks. These findings underscore that reliable continual agent memory systems must account for the complex interactions among accumulated experiences, not just memory scale.