LLM memory is a product feature that stores useful context from earlier interactions so an assistant can behave more consistently over time. Depending on design, it may remember preferences, summaries, task details, or other user-specific state across sessions instead of treating each chat as isolated.
In cyber security, memory matters because it can extend the impact of untrusted input. If a model is exposed to prompt-injected text, poisoned preferences or misleading summaries may persist and influence later answers, recommendations, or automated actions. That makes memory a potential persistence layer, not just a convenience feature.
Defenders should treat memory as controlled data. Keep user intent, retrieved content, and system instructions separate; restrict what may be written to memory; and provide review, deletion, and disable options. In higher-risk environments, non-persistent modes or temporary sessions can reduce the chance that hostile content survives beyond one interaction.