ISOM-R2 streams 1,055,402 tokens at 3.24 GB peak VRAM

The new recurrent memory architecture ISOM-R2 (Isometric State Space / Virtual SVD) ingested a 1,055,402-token context of real Python code on the Aazhi-Coder-1.5B model while holding peak VRAM at 3.24 GB — 8.7x less than the 28.2 GiB a standard KV cache would need. Throughput reached about 8,526 tokens/sec with 100% macro retrieval precision.

• 1,055,402-token context from 181 transformers repo files processed in 123.78 seconds
• Peak VRAM 3.24 GB vs 28.2 GiB for a standard KV cache — 8.7x lower
• Prefill stayed flat at 2.97 GB across all 1M tokens, active buffer under 400 MB
• 100% retrieval precision: only 2 target files isolated from 516 chunks