A new protocol lets AI agents merge what they know, not just what they can do.
MELD, described in a new arXiv paper, gives each agent's memory a five-way decision for every incoming fact: insert it, merge it with an existing entry, link it to a related fact, flag it as a conflict, or reject it. The call is made from three signals - whether the fact already exists under a matching key, how close its embedding is to what's stored, and a natural-language-inference check for whether two statements actually agree. Every accepted change runs through a single auditable "Patch," the only object allowed to alter an agent's memory. A lightweight status CRDT then keeps multiple agents' memories in sync over ordinary publish/subscribe messaging, with no central coordinator required.
Multi-agent AI systems can already call each other's tools through protocols like MCP, but there has been no standard way to reconcile facts two agents phrase differently, or handle it when they flatly disagree. MELD's choice to preserve contradictions rather than silently overwrite them addresses a real failure mode: agents confidently acting on stale or conflicting information because nobody flagged the disagreement. The paper's numbers back the design - on the HotpotQA benchmark, distributed merging matched a centralized memory store's recall while using about 11% less storage than naive merging, and the conflict classifier hit a 0.968 AUC with a 0.013 false-merge rate.
None of the underlying math is new. CRDTs already keep apps like Google Docs and Figma in sync across edits, and version-control systems have handled merge conflicts for decades. What's new here is pointing that machinery at agent memory, a problem multi-agent frameworks have mostly ignored by letting one agent's version win.