Эксперимент: Persistent Relational Memory (#1330)

{ "worked": true, "confidence": 8, "verdict": "Built a reproducible persistent edge-GRU MVP and verified its core math: inactive states are retained exactly, and the gated interpolation keeps each updated coordinate within the old-state/candidate bound. On the synthetic intermittent-pair task, persistent edge memory reduced reactivation MSE from 1.004 to 0.000194 and all-timestep MSE from 0.977 to 0.160, clearly demonstrating the claimed history-recall effect. However, this unoptimized recurrent implementation was about 23.5x slower and used 273 versus 65 parameters, so the accuracy win does not currently satisfy the proposed low-overhead expectation.", "metrics": { "baseline": "Reactivation MSE 1.00384; all-timestep MSE 0.97747; 65 parameters; 0.650 ms per batch of 128", "idea": "Reactivation MSE 0.0001935; all-timestep MSE 0.15959; 273 parameters; 15.256 ms per batch of 128; 0.000193x baseline reactivation error; 23.48x latency" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "result.json", "README.txt" ], "limitations": "This is a small synthetic one-pair-per-sequence recall task, not a full multi-particle dynamic GNN or physical rollout. It does not test node-GRU memory, attention aggregation, eviction/decay, equal-parameter/FLOP comparisons, or generalization across graph structures; the edge dictionary is represented by independent batch slots rather than a production GPU hash map." }