Temporal Memory in a Binary Compressed Automaton

Temporal Memory in a Binary Compressed Automaton (TM-BCA) is an inference-time architecture in which an evolving conversational state directs selective evidence acquisition and bounded reasoning around an unchanged language model. It records current constraints, uncertainty, evidence references and dependencies, detects barriers to a supported decision, and commits a justified result back into temporal state. This paper specifies that architecture and evaluates an implemented deterministic subset. The v0.3 prototype combines aggregate constraint state, incremental dependency updates, a transactional evidence archive, checkpoints, bounded context export and a separate resource-constrained graph planner. Validation passed 65 test methods, 32,496 structured-state comparisons and 3,500 graph-cost comparisons without a mismatch. Against v0.2.1, paired local measurements found approximately 93-fold faster synchronization after one external event, 5.2-fold faster cold reopening and 6-fold faster failed-insert recovery. Repeated large diagnostic payloads used 96.6% less archive space. Short-event archives grew approximately 12-13%, ordinary updates were slightly slower and measured RSS exceeded v0.2.1. Retained Python state was substantially smaller than the historical full-history prototype. These results support the feasibility of compact task-specific state and selective recomputation with recoverable evidence. They do not validate natural-language extraction, calibrated barrier routing, language-model reasoning quality or actual token savings. No real model calls were executed. The architecture has substantial precedents; this paper makes no verified novelty or general superiority claim. Its remaining empirical question is whether the proposed integration improves task quality at a favorable total resource cost on held-out conversational workloads.