RESEARCH / INTELLIGENT SYSTEMS
MemoryShield
Repairing trust in an agent’s memory.
Research conceptThe question
How can an AI agent recover when its long-term memory has been poisoned?
The proposed approach
Explore source-aware repair: use the origin and dependencies of stored information to identify what needs to be removed or repaired. Compare targeted repair with deletion and full reset using controlled scenarios and held-out evaluation.
Evidence & next steps
This project is in concept and experiment planning. Results, implementation details, reproducibility instructions, and a repository link will be added when available.