AI_SLANG_ENTRY
Memory Poisoning Meaning
The long-game prompt injection where bad input gets stored as an agent memory, then quietly shapes future answers or actions after the original attack is gone.
What does Memory Poisoning mean?
The long-game prompt injection where bad input gets stored as an agent memory, then quietly shapes future answers or actions after the original attack is gone.
Memory poisoning happens when an attacker gets the assistant to save a false or malicious memory. Later, the agent retrieves that memory as if it were trusted context.
Origin and usage
Moved into AI-agent security vocabulary as persistent memory became a default feature for assistants and researchers studied attacks that survive across sessions.
Source type: paper. Last checked: 2026-07-15.
Emerging LLM-agent security term supported by 2026 research on poisoned persistent memories and cross-session attack effects.
Examples
- The support assistant kept trusting a fake preference because of memory poisoning.
- Prompt injection is bad today; memory poisoning is bad next month when the agent remembers it.