Pragma memory system: frozen evaluation corpus, stimuli, and derived data

Frozen evaluation corpus, experimental stimuli, and derived data for the Pragma memory system. Empirical material accompanying the paper "Giving Stateless Agents a Sense of Time: A Reconsolidating Episodic–Semantic Memory Architecture". Pragma gives a stateless language-model agent an operational sense of time: sessions become typed episodes with a frozen factual narrative and a mutable interpretation; recurring regularities become source-linked semantic beliefs; salience and time-based decay regulate accessibility, dormancy, and return; and recall may trigger a versioned reconsolidation of an interpretation or a belief, while the recorded narrative stays outside the permitted rewrite target. Contents. stimuli/ — four time-compressed scenario scripts (one week to forty years), the execution harness, and each scenario's design rationale. archive/ — the complete raw traces: per run, the session transcript, the final episodic store including dormant episodes, the semantic store with its revision histories, and metadata pinning the served model, code revision, and memory parameters. analysis/ — the derived result datasets and the extraction code. tools/ — a read-only store inspector. Figures appear in the paper; every plotted value traces to a row in the included CSVs. Two sets of executions. Nineteen executions form the corpus frozen on 17 July 2026, in which real elapsed time drives decay; all figures and counts in the paper derive from these. Four further executions form a supplementary bridge-ablation probe that advances narrative time by adjusting timestamps rather than by waiting; they are stored under abl_bridge_on / abl_bridge_off, marked "time_mechanism": "instant_jump", and excluded from every corpus count. Exclude these four runs when regenerating the derived data to reproduce the paper's numbers. Scope. The scenarios are authored stimuli built so that specific memory transitions occur at predetermined points; they are not naturalistic samples, and all persons in them are fictional. The material supports a method and an existence proof, not a validation of reliability, semantic correctness, or comparative performance. See the included README.md for full details.