MAP-Graph: Provenance-Aware Shared Memory for Multi-Agent Workflows

A recent study presents MAP-Graph, a memory layer that is aware of provenance and is intended for managing shared memory within multi-agent language model processes. This system tackles the issue of ensuring that pertinent evidence is valid for particular agents or actions, as limitations can spread through derivations, resulting in unauthorized reads or unsafe behaviors. MAP-Graph depicts agents, sources, memories, claims, and actions within a typed execution graph, tracing lineage and eliminating records that lack permission. It reorders eligible memories based on semantic similarity and multiplicative path trust, implementing a risk-sensitive gate prior to action execution while preserving impacted lineage for auditing purposes. The findings, detailed in a paper on arXiv (2608.10509), highlight the system's efficiency across a controlled benchmark of 2,700 tasks.