HaluMem
Exposing Memory Hallucinations in AI Systems
The Memory Crisis in AI
Modern AI systems are experiencing a critical memory crisis. When you tell a chatbot you enjoy science fiction, it might later recommend romance novels. Share news about your promotion, and minutes later it asks about your job.
This isn't just forgetfulness—it's memory hallucination, a fundamental flaw where AI systems fabricate, misremember, and generate contradictory content, contaminating all subsequent reasoning.
Four Types of Memory Hallucinations
Fabrication
Creating memories that never happened
User: "I like parrots now" → System stores: "User dislikes parrots"Error
Wrong details in retrieved memories
User: "My friend Joseph..." → System remembers: "Friend Mark..."Conflict
Contradictory memories coexist
Memory 1: "Good health" + Memory 2: "Poor health" (both active)Omission
Failing to retrieve crucial information
User mentions promotion → System forgets it in next conversationCore Memory Operations
Extract
Identify key information from conversation
Store
Save as structured memory points
Update
Modify when encountering new info
Retrieve
Fetch relevant memories for context
Max Recall Rate
Best systems only retrieve less than half of important memories
Max Accuracy
Large portion of stored memories are fabricated or incorrect
Update Omission
Systems fail to update memories when new info arrives