Title: Rashomon Memory: Enabling Argumentation-Driven Retrieval for Multi-Perspective Agent Memory

Abstract: As AI agents operate across long temporal spans, they gather experiences that must simultaneously support various objectives, often necessitating the maintenance of contradictory interpretations of identical events. For instance, a concession made during client negotiations might be encoded as a "trust-building investment" to advance one strategic objective, while being viewed as a "contractual liability" for a different goal. Existing memory systems typically presuppose a single correct encoding or, at most, allow for multiple views within a unified storage structure. In contrast, we introduce Rashomon Memory, an architectural framework in which parallel, goal-conditioned agents encode experiences based on their specific priorities. These agents engage in argumentation during query time to negotiate outcomes. Each perspective retains its distinct ontology and knowledge graph. During the retrieval phase, these perspectives put forward interpretations and critique one another’s proposals using asymmetric domain knowledge, with Dung’s argumentation semantics deciding which proposals endure. The resulting attack graph serves as an explanation in itself, documenting the selected interpretation, the alternatives evaluated, and the reasons for their rejection. Our proof-of-concept demonstrates that retrieval modes—such as selection, composition, and conflict surfacing—emerge from the topology of the attack graph. Notably, the conflict surfacing mode allows the system to report genuine disagreements rather than enforcing resolution, thereby enabling decision-makers to observe the underlying interpretive conflicts directly.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC