Researchers have built a memory system that lets AI agents borrow experience from each other without getting confused by the details of a brand-new environment.
The system, called MemCo, splits an agent's memory into two layers. A local memory keeps environment-specific details, like the exact steps that worked in a particular app or game. A global memory extracts transferable workflows from those local trajectories, stripping out details that would not make sense elsewhere. When an agent faces a new situation, MemCo checks its current state and decision stage, then pulls in a mix of local and global memories suited to that moment, rather than retrieving one giant, undifferentiated memory. The researchers tested the approach on interactive decision-making benchmarks and found it improved task success and cut down on wasted exploration compared to agents using either isolated memory or a single shared memory pool.
This is a practical fix for a real bottleneck. Collecting enough trajectories to make an agent's memory useful is expensive, and dumping every agent's experience into one shared pool creates a mismatch problem: memories end up either too specific to a prior task to be useful, or too vague to guide the next action. By routing memory at the right granularity, MemCo lets agents share know-how without inheriting each other's blind spots.
It is a reminder that scaling agent memory is less about hoarding more data and more about organizing what is already there.