Streaming Belief Propagation on Mixed-Alphabet Tanner Graphs for Practical Quantum Memory

Reliable quantum memory under circuit-level noise requires decoders that can process syndrome information continuously and at a rate comparable to its generation. In practical quantum error correction (QEC), repeated syndrome measurements cause the number of potential error locations to grow rapidly with both code size and time. In this paper, we propose a streaming mixed-alphabet belief propagation (SM-BP) decoder. We construct a space-time Tanner graph across multiple rounds of syndrome extraction with mixed-alphabet error variables, preserving correlations arising from multi-qubit faults. Additionally, we propose an adaptive sliding window procedure that captures long error events across window boundaries and adjusts the decoding in real time. To enhance SM-BP, we introduce a technique of probabilistic error consolidation to mitigate degeneracy effects and short cycles. Our simulations demonstrate high error thresholds of 0.4% to 0.87% and strong error-floor performance for topological code families, including rotated toric, toric color, and twisted XZZX toric codes. These results show that SM-BP provides a practical decoding framework for continuous QEC under circuit-level noise.