Independent reproduction wanted: ASM-CM 32K associative recall

[gnai-creator]

ASM-CM is an experimental compact-memory architecture designed to process long streams while keeping the retained neural state approximately bounded.

In the current frozen protocol, the model achieved:

• 100% MQAR associative recall at 32K;
• confirmation across three independent seeds;
• approximately 140 KiB of retained tensor state per stream;
• no proportional growth of retained state as the evaluated stream length increased.

This does not mean infinite memory.

The model retains a finite compact state and may forget, suffer interference, or fail under harder distributions and longer streams. The current result is limited to the evaluated protocol and should not be interpreted as proof of unlimited context retention.

I am looking for independent researchers and engineers willing to reproduce, stress-test, or falsify the result.

Useful experiments include:

• contexts longer than 32K;
• additional seeds;
• different GPUs and software environments;
• denser key-value assignments;
• adversarial interference between similar keys;
• repeated overwriting of associations;
• delayed retrieval at different distances;
• measurements of forgetting and failure thresholds;
• comparisons with Transformer, RNN, SSM, and retrieval-based baselines;
• verification of retained state, throughput, latency, and peak VRAM.

Negative results are welcome.

If the result fails under a configuration, please report:

• the exact commit;
• command used;
• configuration file;
• hardware and software environment;
• random seed;
• observed metrics;
• logs or artifacts needed for reproduction.

The goal is not to defend the architecture from failure. The goal is to determine precisely where it works, where it stops working, and whether the reported behavior can be reproduced independently.

Repository and reproduction instructions:

https://github.com/gnai-creator/ASM

Please use this discussion for questions, reproduction reports, failed runs, protocol critiques, and proposed stress tests.