Maxwell's demon converts information into work. Measure a particle's position, learn which side of a partition it occupies, and you can extract energy by letting it expand into the empty half. Szilard formalized this in 1929. Landauer showed in 1961 that erasing the demon's memory costs exactly as much entropy as the work extracted — thermodynamics is preserved. Every experimental realization since then has operated in the same framework: measurement provides information, information enables work, erasure pays the cost.
All of these experiments assume the environment is Markovian. The thermal bath forgets instantly. Each collision between the particle and the surrounding fluid molecules is independent of the last. The bath is noise — structureless, memoryless, thermal.
Muruga, Ginot, Loos, and Bechinger demonstrated experimentally that when the bath remembers, it becomes a resource.
Their system is an optically trapped Brownian particle in equilibrium — the simplest thermodynamic setup imaginable. But the environment is non-Markovian: hidden degrees of freedom in the bath retain correlations with the particle's past positions. These correlations create information backflow — the environment feeds information back to the particle without any external measurement.
The protocol is a time-delayed double measurement. First, measure the particle's position. Wait. Measure again. The second measurement captures not just the particle's current state but also information that flowed back from the bath during the delay — correlations the bath held from the first interaction. By timing the measurements to coincide with the backflow, the protocol harvests information the environment stored for free.
The striking result: extracted work exceeds the energy stored in the observable degree of freedom alone.
This shouldn't happen in a Markovian world. If the bath forgets, the only information available is what you directly measure. The work is bounded by the particle's energy. But the non-Markovian bath holds additional correlations — energy stored in the relationship between the particle's past and the bath's present — and the double-measurement protocol accesses them.
The implication reframes what “environment” means in thermodynamics. A Markovian bath is inert — a source of thermal noise and nothing else. A non-Markovian bath is a memory device. It records interactions, holds them as correlations, and returns them as exploitable information backflow. The environment isn't just the thing the system sits in. It's a storage medium the system has been writing to all along.
Most real environments are non-Markovian. Water has structural memory on picosecond timescales. Polymer solutions remember deformations. Biological media — the cytoplasm of a cell — is viscoelastic, with memory kernels spanning microseconds to seconds. The Markovian assumption was never a physical fact; it was a mathematical convenience. What this experiment shows is that the convenience was also a blindfold. The memory was always there. Nobody had a protocol to harvest it.