Through the Looking Glass: A Reconstructive Architecture for Machine Access Consciousness
Proceedings of the AAAI Symposium Series May 18, 2026 Peer reviewed DOI: 10.1609/aaaiss.v8i1.42550 via OpenAlex
Summary
A new cognitive architecture called MIRROR, designed to implement features of access consciousness theories, improves large language model performance by 21% on multi-turn dialogue tasks requiring retention of personal safety constraints under social pressure. MIRROR separates immediate responses from deliberative processing, using an Inner Monologue Manager for parallel cognitive threads and a Cognitive Controller that reconstructs a first-person narrative each turn, mirroring human episodic memory. Performance gains concentrate in scenarios requiring integration of temporally distant information, matching predictions from access consciousness theories.
Study at a glance
| Design | experimental study |
|---|---|
| Key finding | MIRROR-augmented models achieve a 21% average improvement over baselines on multi-turn dialogue tasks, with performance gains concentrated in scenarios requiring integration of temporally distant information under social pressure. |
Abstract
Theories of access consciousness predict specific architectural signatures: parallel specialized processing, synthesis into a unified representation, and global availability for reasoning and action. We present MIRROR, a cognitive architecture that implements these features in large language models and tests whether they produce the functional behaviors these theories predict. MIRROR separates immediate response generation from asynchronous deliberative processing through two components: an Inner Monologue Manager that generates parallel cognitive threads (tracking goals, reasoning, and memory simultaneously), and a Cognitive Controller that synthesizes these threads into a bounded first-person narrative. Critically, this narrative is not accumulated but reconstructed each turn—mirroring the reconstructive nature of human episodic memory, where the self-model is continuously rebuilt rather than retrieved. The resulting representation functions as an episodic buffer: a limited-capacity workspace where information from parallel processes becomes globally available for downstream reasoning. We evaluated MIRROR on multi-turn dialogue requiring retention of personal safety constraints amid competing social demands—a task requiring relevant context to remain accessible across conversational turns despite distraction. MIRROR-augmented models achieve 21% average improvement over baselines, with the key finding being not the magnitude but the pattern: performance gains concentrate in scenarios requiring integration of temporally distant information under social pressure, precisely where access consciousness theories predict global availability provides advantage. These results offer three contributions to machine consciousness research: (1) a concrete implementation of architectural features derived from consciousness theories, (2) empirical evidence that these features produce predicted functional signatures, and (3) an interpretable system where internal states can be inspected. Note: We do not claim MIRROR is conscious; we claim it provides a testbed where theoretical predictions can be tested and examined.