Clark and Chalmers (1998) proposed that cognitive processes can extend beyond the brain when external artefacts function as reliable parts of an agent’s cognitive routine. Twenty-seven years later, generative AI offers qualitatively new possibilities for such extension. Clark (2025) has argued that large language models represent a genuinely novel form of hybrid thinking system. Barandiaran and colleagues (2025) have introduced the concept of “midtended cognition” to describe the distinctive agency that emerges when humans couple with generative AI. Empirical accounts of living with such systems are only beginning to appear: a six-month analytic autoethnography documents the relational labour of sustaining user-built continuity infrastructure (Jiang, 2026), and a formal framework now specifies criteria under which coupled AI systems constitute rather than merely model cognition (Zhao, 2026). Yet no study has built a persistent, bidirectional extended cognitive system, inhabited it over a sustained period, and benchmarked its effects on the operator’s cognition. This paper reports the design, construction, and empirical evaluation of the Resonator: a personally owned cognitive externalisation stack comprising persistent vector memory, context compilation, live interaction with a large language model, session decompilation, and memory write-back. The system closes a bidirectional loop—compiler to session to decompiler to memory to compiler—that maintains cognitive continuity across sessions without requiring the operator to manually reconstruct context. A controlled within-subjects experiment (N = 16 matched prompts across four knowledge domains) compared three conditions: cold-start baseline, compiler-primed, and full-loop (compiler plus decompiler write-back). Wilcoxon signed-rank tests yielded significant improvements at each comparison (Stage 1 vs Stage 2: p = 0.0019, r = 0.574; Stage 1 vs Stage 3: p = 0.0009, r = 0.772; Stage 2 vs Stage 3: p = 0.0039, r = 0.331). Token consumption dropped 75–84% in compiler-primed sessions relative to cold starts. A pre-registered goal-alignment test demonstrated that the system produced output shaped by strategic goals never stated in the active session but held in the memory substrate, confirming that cognitive coordination was offloaded to the external store. Four documented instances of the system detecting and correcting errors in its own substrate provide evidence of metacognitive-analogue self-repair. The paper argues that these findings satisfy a strengthened version of Clark’s parity principle: the external system not only performs functions equivalent to internal cognition but also exhibits governance-level self-correction—detecting and repairing its own substrate faults without external intervention. The evidence additionally bears on the bidirectionality criterion of Zhao’s (2026) externalisation framework, while diverging from that framework’s substitutive trajectory: the system described here is designed for augmentation, not replacement. Keywords: extended mind, cognitive extension, human–AI coupling, parity principle, personal knowledge management, autoethnography, bidirectional memory, context compilation.