MemFuse: Multi-Source Memory Fusion from Fragmented Observations

TL;DR - MemFuse introduces a benchmark and structured memory system for agents that must combine fragmented, source-tagged observations into coherent long-term memories. It matters because realistic agent memory requires cross-source reasoning without losing evidence provenance.

• MemFuseBench uses a Scene-to-Sensor pipeline to generate controllable observations, evidence-grounded questions, adversarial distractors, and source metadata.
• The benchmark evaluates temporal reasoning, cross-source evidence fusion, and robustness to noise.
• MemFuse stores source-level evidence as atomic events, then organizes related events into fused memories through a causal fusion graph.
• Across three LLM settings, MemFuse achieved the best overall performance among the evaluated memory systems and consistently improved cross-source fusion tasks.