StateFuse Offers Conflict-Preserving Memory for Multi-Agent Systems

In multi-agent systems, agents often generate conflicting observations from various sources like parallel branches or retries. Current memory systems frequently resolve these conflicts by simply overwriting older data, making it difficult to understand or correct past disagreements. StateFuse addresses this by providing a new memory contract that explicitly preserves conflicts. StateFuse is built upon standard OpSet/CRDT merge principles but adds an agent-facing semantic layer. This layer ensures an immutable history, clearly identifies conflict objects, and provides precise handles for correction. Crucially, it resolves conflicts at the projection stage without altering the underlying replicated state, ensuring auditability. Evaluations against traditional memory approaches showed that while all methods achieved similar answer accuracy, StateFuse's ability to keep contradictions visible allowed for safer decision-making, such as abstaining when conflicts are present, and more auditable corrections. This makes StateFuse particularly valuable for applications where transparency, safety, and the ability to correct past errors are paramount.