Researchers just automated one of the fiddliest parts of building AI agents: giving them a memory that actually works.
A new system called AutoMem searches through hundreds of possible memory setups - combinations of 5 encoders, 5 stores, 6 retrievers, and 4 managers - to find which one fits a given task best. The researchers first showed why this matters: no single memory design wins across the board, and picking the wrong one creates real performance gaps between tasks. AutoMem closes that gap with two feedback loops. One proposes new architectures based on what has worked in past search runs; the other traces failures back to the specific module that caused them and turns that into targeted feedback the system can act on.
Tested on three agent benchmarks - GAIA, WebWalkerQA, and xBench-DeepSearch - across two different underlying models, AutoMem beat the best hand-tuned memory setups by 2.8 accuracy points on average. On one backbone it also cut token costs by 14.3 percent. That second number matters as much as the accuracy gain: most agent upgrades add compute cost rather than save it.
It is a narrow, technical win rather than a leap - but as agent memory keeps getting bolted on ad hoc, a system that tunes itself instead of shipping with a one-size-fits-all default is worth watching.