Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory

Researchers propose Rashomon Memory, a new AI agent memory architecture where multiple goal-conditioned agents maintain parallel interpretations of the same events and negotiate through argumentation at query time. The system allows AI agents to handle conflicting perspectives on experiences rather than forcing a single interpretation, using Dung's argumentation semantics to determine which proposals survive retrieval.

• →Current AI memory systems assume single correct encodings, but real-world agents need to maintain conflicting interpretations of the same events for different goals.
• →Rashomon Memory uses parallel goal-conditioned agents that each maintain their own ontology and knowledge graph for encoding experiences.
• →At retrieval time, different perspectives propose interpretations and critique each other using asymmetric domain knowledge.
• →The system generates attack graphs that serve as explanations, showing which interpretation was selected and why alternatives were rejected.
• →Three retrieval modes emerge: selection, composition, and conflict surfacing, with the latter allowing decision-makers to see underlying interpretive conflicts directly.