Formal foundations of the memory-system account of the epistemic arrow of time: taxonomy completeness, a Type-3 memory-of-future no-go, the cosmological turnover magnitude, and the inductive footing of memory

The Wolpert–Kipper framework ('Memory Systems, the Epistemic Arrow of Time, and the Second Law', Entropy 2024) gives a fully formal account of why memory and records inform us only about the past: among three types of memory system — physical systems whose present state carries mutual information about the external world at another time, quantified by restricted mutual information — only Type-3 systems (records with an entropy-increasing initialized state, such as photographs, footprints, and arguably human memory) are time-asymmetric, exploiting the second law of thermodynamics. The authors flag four issues requiring formal resolution: quantify the Brownian-bridge 'turnover' effect for the entropy of the actual universe (conditioning a time-symmetric Markov entropy process on its values at both the Big Bang and the present can raise the expected recent-past entropy above the present value); prove or refute the completeness of the three-type taxonomy; prove or refute that no Type-3 system can provide reliable information about the future; and resolve how the joint distributions the definitions presuppose can be known without circular appeal to memory. These four issues remain open.

Background

The epistemic arrow of time is the observation that our knowledge of the past is both of a different kind and more detailed than our knowledge of the future: we have memories and records of the past, but nothing analogous for the future. A widespread explanatory strategy in the foundations of statistical mechanics grounds this asymmetry in the second law of thermodynamics — the tendency of the coarse-grained entropy of isolated systems to increase — which is itself standardly derived from the Past Hypothesis: the posit that the very early universe was in a state of extraordinarily low entropy.

Wolpert and Kipper (building on Wolpert's earlier semi-formal 1992 analysis) gave this strategy its first fully formal framework. A memory system is a physical system $M$ whose present state can carry information about the state of the external world $W$ at some time other than the present. Writing $m_t$ and $w_t$ for the states at time $t$ , the framework models the joint dynamics of $M \times W$ as a stochastic process and measures memory content by restricted mutual information: the mutual information between $W_1$ (the world state at a past time $t_1$ ) and $M_0$ (the memory state at the present $t_0$ ), evaluated under the conditional distribution given that $m_0$ lies in a specified set $M^*$ (or $w_0$ in a set $W^*$ ). Three types of memory system are identified in our universe. Type-1 systems condition on $m_0 \in M^*$ ; Type-2 systems (exemplified by engineered computers, which almost deterministically compute the joint dynamics) condition on $w_0 \in W^*$ ; both are time-symmetric, able to inform about the future as well as the past. Type-3 systems are the special case of Type-1 that formalizes records such as photographs, footprints on a beach, and impact craters: there is an initialized (ready) state $m^\dagger$ such that an entropy-increasing physical process — exposure of film, a foot pressing sand, a wiping circuit coupling the memory to an external initializing system $K$ — reliably prepares $M$ in $m^\dagger$ at a past time $t_1$ . With the ordering $t_2 < t_1 < t_0$ , the definition's conditions (2a) and (2b) state that, given the present record $m_0 \in M^*$ : (2a) the pre-initialization state $m_2$ lay in an unremarkable set $M^\prime$ with probability close to 1, and (2b) the state $m_1$ at $t_1$ is retrodictable as sharply peaked on $m^\dagger$ . This is what licenses using the present record as information about the world state $w_1$ at $t_1$ , and it is precisely this two-step structure that relies on the second law in every known instance — which is why Type-3 memories inform only about the past. The authors argue that human memory is Type-3, which would complete the derivation of the epistemic arrow from the second law.

Two further concepts are needed for the open issues. First, the Brownian-bridge (turnover) problem: the Past Hypothesis is standardly formalized by modeling the entropy $S_t$ of the universe as a time-symmetric Markov process — a Fokker–Planck diffusion or a master equation, the standard continuous-time and discrete-state stochastic models for the evolution of a probability distribution over states — and conditioning on a low early value. But we know the present entropy at least as reliably as the early one, and correct Bayesian conditioning on the values at two times yields a Brownian bridge (a Markov process conditioned on both its endpoint values): the conditional expected entropy between the two times can rise above both endpoint values, and its time derivative at the present can be negative. In mesoscopic systems this turnover is observable in principle; whether it is negligible at cosmological scales is undetermined. Second, the inductive footing: the restricted-mutual-information definitions presuppose that the memory's user knows the joint distribution $P(m_0, m_1, w_0, w_1, \ldots)$ , and it is unclear how that knowledge could be acquired without already relying on memory — a confrontation with the formal impossibility of inductive inference (no-free-lunch-type results). The source paper's closing section flags the turnover computation, the taxonomy-completeness question, the missing future no-go for Type-3 systems, and the inductive-justification problem as issues requiring future work. The paper also remarks, descriptively, that the initializing system $K$ does not explicitly occur in the Type-3 definition but 'arises in practice, as part of the underlying process that enforces' condition (2a).

Problem Statement

Resolve, or refute the adequacy of, the formal foundations of the Wolpert–Kipper account of the epistemic arrow of time, by discharging the four gaps flagged in the source paper's Section 6:

(Cosmological turnover.) Compute the two-time-conditioned entropy dynamics $E[S_t \mid S(t_0), S(t_\ast)]$ of a physically motivated Markov model (Fokker–Planck or master-equation, with drift and diffusion terms fixed by cosmology) of the coarse-grained entropy of our universe, where $t_0$ is the present and $t_\ast$ the time of the Big Bang; and determine whether the bridge-induced elevation of the expected $S_t$ in the recent past, or a negative time derivative of the conditional expectation at the present, is negligible at cosmological scales or materially alters standard derivations of the second law from the Past Hypothesis. (The formal two-time framework and two toy Markov models — a random walk and an idealized gas — already exist; the cosmologically realistic computation does not.)
(Taxonomy completeness.) Prove, or construct a counterexample to, the claim that within physical theory the Type-1/Type-2 classification (with Type-3 as the special record-forming case) exhausts the ways a system's present state can carry large restricted mutual information about the world at another time; in particular, determine whether a physically realizable mechanism can robustly produce memory-of-future instances not captured by the Type-1/Type-2 formalism.
(Type-3 future no-go.) Prove or refute: no Type-3 memory system can, even in principle, provide reliable information about the future state of the external world. The source paper offers plausibility arguments only — no formal impossibility proof exists.
(Inductive footing.) Establish how the user of a memory system can come to know, or justifiably assign, the joint distributions $P(m_0, m_1, w_0, w_1, \ldots)$ presupposed by the restricted-mutual-information definitions, without circular appeal to the reliability of prior memory — the framework's confrontation with the formal impossibility of inductive inference. A treatment discharges this item only by giving an explicit, non-circular account — a formal construction or a defended epistemic argument — of how that knowledge or justified assignment is obtained, rather than merely asserting that such assignment is possible in principle.

The four components are jointly the completion of one objective and are not independent: item 1 secures the physical premise — the second law as it actually follows from the Past Hypothesis under correct two-time conditioning — on which the Type-3 asymmetry exploited in item 3 rests; items 2 and 3 together would establish that the past-directedness of memory is forced by the framework rather than accidental; and item 4 secures the epistemic license to use the distributions invoked in items 2 and 3. A resolution is a treatment that discharges all four gaps for the formalism as stated, or demonstrates precisely where and why the account as a whole fails. Refuting any one component resolves that component alone and does not by itself resolve the task; and a demonstration that the account as a whole fails must name the component on which that failure turns.

References

David H. Wolpert and Jens Kipper, 'Memory Systems, the Epistemic Arrow of Time, and the Second Law', Entropy 26(2):170 (2024). DOI: 10.3390/e26020170, https://doi.org/10.3390/e26020170 (arXiv:2309.10927, https://arxiv.org/abs/2309.10927).
Jordan Scharnhorst, David Wolpert, and Carlo Rovelli, 'Boltzmann Bridges' (2024). arXiv:2407.02840, https://arxiv.org/abs/2407.02840. (arXiv v2 of 7 August 2024.)
David H. Wolpert, Carlo Rovelli, and Jordan Scharnhorst, 'Disentangling Boltzmann Brains, the Time-Asymmetry of Memory, and the Second Law', Entropy 27(12):1227 (2025). DOI: 10.3390/e27121227, https://doi.org/10.3390/e27121227 (arXiv:2507.10959, https://arxiv.org/abs/2507.10959).
Carlo Rovelli and David Wolpert, 'Why you do not need to worry about the standard argument that you are a Boltzmann brain' (2024). arXiv:2407.13197, https://arxiv.org/abs/2407.13197.
Aram Ebtekar and Marcus Hutter, 'Modeling the Arrows of Time with Causal Multibaker Maps', Entropy 26(9):776 (2024). DOI: 10.3390/e26090776, https://doi.org/10.3390/e26090776.
Luis Pedro García-Pintos, Yi-Kai Liu, and Alexey V. Gorshkov, 'Reshaping the Quantum Arrow of Time', Physical Review X 16, 011028 (2026). DOI: 10.1103/l18s-9vmh, https://doi.org/10.1103/l18s-9vmh.
David H. Wolpert, 'Memory systems, computation, and the second law of thermodynamics', International Journal of Theoretical Physics 31(4), 743–785 (1992). DOI: 10.1007/BF00670829, https://doi.org/10.1007/BF00670829.
Jordan Scharnhorst and Anthony Aguirre, 'Modeling the Past Hypothesis: A Mechanical Cosmology', Foundations of Physics 54, article 8 (2024; published online 19 December 2023). DOI: 10.1007/s10701-023-00745-3, https://doi.org/10.1007/s10701-023-00745-3.
Chas A. Egan and Charles H. Lineweaver, 'A larger estimate of the entropy of the universe', The Astrophysical Journal 710, 1825–1844 (2010). DOI: 10.1088/0004-637X/710/2/1825, https://doi.org/10.1088/0004-637X/710/2/1825.
Meir Shimon, 'Rapidity Asymmetry in Cosmology and the Observable Cosmological Arrow', Universe 12(6):182 (2026). DOI: 10.3390/universe12060182, https://doi.org/10.3390/universe12060182.