Characterizing perfect recall using next-step temporal operators in S5 and sub-S5 Epistemic Temporal Logic

We review the notion of perfect recall in the literature on interpreted systems, game theory, and epistemic logic. In the context of Epistemic Temporal Logic (ETL), we give a (to our knowledge) novel frame condition for perfect recall, which is local and can straightforwardly be translated to a defining formula in a language that only has next-step temporal operators. This frame condition also gives rise to a complete axiomatization for S5 ETL frames with perfect recall. We then consider how to extend and consolidate the notion of perfect recall in sub-S5 settings, where the various notions discussed are no longer equivalent.

💡 Research Summary

The paper revisits the notion of perfect recall—a property stating that an agent never forgets information it once knew—across three research traditions: interpreted systems, game theory, and epistemic logic. In interpreted systems, perfect recall is usually expressed as a global condition on histories; in game theory it guarantees that strategies can be based on the full past; in epistemic logic it appears as “history‑consistency” or “information‑preservation.” All of these formulations are inherently non‑local: they require quantification over entire runs or histories, which makes both theoretical analysis and automated verification cumbersome.

The authors’ main contribution is a novel, purely local frame condition that captures perfect recall using only the next‑step temporal operator X and the standard epistemic operator K_i. The condition can be informally stated as: for any state s and any agent i, whatever i knows at s must still be known after a single transition to the successor state. Formally, the condition is expressed by the schema
  K_i p → X K_i p
or equivalently by the axiom X‑K‑preserve: X K_i φ ↔ K_i X φ for all formulas φ. Crucially, this formulation does not require the use of global modalities (□, ◇) or any higher‑order temporal constructs; it lives entirely within the next‑step fragment of ETL.

The paper shows that, on S5 frames (where the epistemic accessibility relation is an equivalence relation), the X‑K‑preserve condition is both necessary and sufficient for perfect recall. The authors prove two directions: (1) any S5 model satisfying perfect recall validates X‑K‑preserve, and (2) any S5 model validating X‑K‑preserve automatically satisfies the standard history‑consistency definition. Building on this equivalence, they present a complete axiomatization for S5 ETL with perfect recall: the usual S5 axioms (K, T, 4, 5), the standard next‑step axioms for X, and the additional X‑K‑preserve axiom. Soundness and completeness are established by constructing canonical models that respect both the epistemic equivalence and the next‑step preservation property.

After establishing the S5 case, the authors turn to sub‑S5 settings—frames that lack one or more of reflexivity, symmetry, or transitivity (e.g., K, KD45, T). In these weaker logics the three classic definitions of perfect recall (history‑consistency, information‑preservation, and the new X‑based condition) diverge. The paper provides concrete counter‑examples showing that X‑K‑preserve may hold while an agent still loses some past knowledge, or vice versa, depending on the missing frame property. To reconcile these discrepancies, two supplementary constraints are introduced: (i) past‑accessibility preservation, which requires that the epistemic relation at a predecessor state be included in the relation at its successor, and (ii) possibility‑consistency, which ensures that all worlds considered possible after a transition retain the agent’s prior knowledge. By adding appropriate combinations of these constraints to the base logic, the authors obtain sound and complete axiomatizations for perfect recall in each sub‑S5 family.

Beyond the theoretical results, the paper highlights practical implications. Because X‑K‑preserve is a local, step‑wise condition, it can be checked efficiently by model‑checking tools that operate on transition systems, avoiding the exponential blow‑up associated with global history quantification. This makes the approach attractive for verification of multi‑agent protocols, dynamic game strategies, and knowledge‑based system designs where perfect recall is a desired safety property. Moreover, the analysis of sub‑S5 logics clarifies how designers can deliberately weaken epistemic assumptions (e.g., dropping symmetry to model asymmetric information) while still preserving a suitable notion of recall, by selecting the appropriate auxiliary axioms.

In summary, the paper delivers a novel, next‑step‑only characterization of perfect recall, proves its adequacy for S5 ETL frames, supplies complete axiomatizations for both S5 and weaker epistemic logics, and demonstrates that the local condition dramatically simplifies verification while preserving the essential epistemic intuition behind perfect recall.