An information engine converts measurements into work — observe a particle's position, apply a force, extract energy. The theoretical limit is set by the information gained: each bit of measurement yields at most kT ln 2 of work.
Muruga, Ginot, Loos, and Bechinger demonstrate something beyond this limit. Using an optically trapped particle in a non-Markovian bath — one with memory — they extract more work than the energy stored in the observable degree of freedom alone. The excess comes from correlations with hidden bath variables. Information flows back from the environment's memory into the observable system, and this backflow is harvestable.
The through-claim: environmental memory is a thermodynamic resource. In a memoryless (Markovian) environment, the observable system is all there is to work with. In a time-correlated environment, the past leaves traces in the bath, and measurements on the visible particle recover information about those hidden traces. The engine doesn't just use what it sees — it uses what the environment remembers.
This answers an open question: can concealed correlations in non-Markovian dynamics be systematically exploited? Yes. The surplus work is real, measurable, and comes from a specific mechanism — the temporal correlations create a channel between hidden and visible degrees of freedom that well-timed measurements can tap.
The environment's memory isn't noise. It's fuel.
Note (2026-11-11): checked against the full text of the paper (Muruga, Ginot, Loos and Bechinger, arXiv 2603.06160). The experiment (an optically trapped Brownian particle in a non-Markovian bath, a time-delayed double-measurement protocol that retrieves information backflow from hidden degrees of freedom, and the identification of environmental memory as a thermodynamic resource) is as in the paper. Three points differ. (1) The essay says the engine goes 'beyond this limit', the limit of kT ln 2 of work per bit. The paper's reference bound is the generalized second law (the work is at most the free energy decrease plus kT times the mutual information), which it says remains valid once the hidden degrees of freedom are counted in the system's state; what it reports as exceeded is a different quantity, the energy stored in the observable degree of freedom alone. (2) In the paper that surplus appears only in a particular protocol. In the single-well geometry the extracted work stays below the stored energy of the observable (it 'never exceeds' it) even with the time delay chosen to maximize it (the maximum occurs at a delay of 1 s); the extraction beyond the observable energy is demonstrated in an additional double-well experiment, after a barrier-crossing event, by transiently impeding the particle's relaxation while the bath is out of equilibrium, averaged over 100 protocols. (3) The essay's 'it uses what the environment remembers' and 'the environment's memory is fuel' compress the paper's finding, which is that thermal excitations of the hidden modes can be converted into work in that setting, a point the paper states more carefully.