MAP-Graph Enhances Multi-Agent Memory with Provenance and Trust

Researchers have introduced MAP-Graph, a new provenance-aware shared memory layer designed for multi-agent language model workflows. While shared memory is crucial for agents to reuse information across long workflows, existing systems often struggle with ensuring that relevant evidence is admissible for specific agents or actions, especially when dealing with private, poisoned, untrusted, or revoked sources. This can lead to unauthorized reads or unsafe actions. MAP-Graph addresses these issues by representing all components—agents, sources, memories, claims, and actions—within a typed execution graph. This graph allows the system to trace the ancestry of information, exclude records that are permission-ineligible, and rerank eligible memories based on semantic similarity and a multiplicative path trust score. Crucially, it applies a risk-sensitive gate before action execution, ensuring that actions are only taken with appropriately trusted and authorized information, while retaining the full lineage for auditing purposes. Evaluations on a controlled benchmark demonstrated that MAP-Graph achieved 94.96% overall task success and significantly improved exact decision accuracy and access control compared to other methods. The results highlight the effectiveness of using provenance as an operational control signal, rather than just post-hoc audit metadata, within complex multi-agent environments.