Question
Can earlier inference steps provide useful candidate updates for a predictive-coding model?
Approach
An associative memory uses local context to retrieve stored state deltas. Those deltas can propose changes during the model's error-optimization iterations.
The work explores how this retrieval interacts with reconstruction and classification.
Limits
The project records a tradeoff between those two objectives. Clearer ablations are still needed to isolate the contribution of memory retrieval.
Confidence gating and broader parameter sweeps remain proposed follow-up work. They are not results of the present experiment.