Recovering useful patterns from stored memory

The question

Can the useful pattern combinations found with clean features also be selected from retained bundled memory, without access to the original image or features?

Why this experiment?

A usable memory operation must work from what was stored. Privileged original features are useful assessment references but cannot supply the answer during recall.

How it works

Select patterns using only retained VSA, position roles and the acquired bank. Compare background correction and joint selection, with shuffled memory as a content-specificity control.

What happened

Background correction reduced mean regional squared RGB error by 41.2% versus raw independent selection on the new requests: 0.000266 versus 0.000453. Fitting the pair jointly increased error by 6.1% versus corrected independent selection, to 0.000282. The useful change in this comparison was the background correction.

Lower averages did not establish reliable edits. Centered improved 29 requests, worsened 13 and left 22 unchanged. Joint improved 29, worsened 14 and left 21 unchanged. Among Joint’s 43 changed outputs, 13 increased regional error by more than 10% and ten at least doubled it. Centered had 13 such increases and nine doublings among 42 changed outputs. Both missed the predefined diagnostic benchmark.

The privileged Clean control met all six diagnostic criteria again: mean error 0.000119, with 32 requests improved, one worsened and 31 unchanged. Its one worsening exceeded 10% but did not double error. The same vocabulary and decoder can support useful combinations with clean information; recovering useful combinations from the bundle remains unresolved.

Matching memory content mattered: Joint beat selection from another photograph’s memory on 50 requests, lost on seven and tied on seven. The paired source-level comparison supported that advantage, but not an advantage over Centered. Joint changed only six selected pairs relative to Centered; one improved decoded error and five worsened it. These request-level counts are post-experiment descriptions, not extra acceptance criteria.

Independent review checked 2,777 evidence files and recomputed all 1,536 pixel and edge measurements. Unchanged readbacks and pixels outside the selected core matched exactly. The research life grew from 159 to 175 encounters while preserving earlier memories, the acquired bank, codec and held work. The separate reused panel showed the same direction of results and was not pooled with the new photographs.

Limits and next question

Image, region and feature-position addresses were supplied. This tests a represented feature exchange, not autonomous object discovery or object motion. The successful Clean selector remains privileged, and automatic constituent substitution stays disabled.

Large proportional error increases can concern small requested effects; they do not mean the whole image was destroyed. The diagnostic criteria retain those regressions rather than allowing large gains elsewhere to cancel them. Two weak-reference requests remain in the 64-request totals and contribute zero normalized gain.

The estimated background is the learned bank’s mean contribution, not the actual contents of the other positions. Failure of this recovery rule does not establish that the retained information is absent or that another recovery method would fail. The clean reference is not an upper bound on decoded quality.

Next question. Can a shared learned cleanup component estimate the feature difference needed for an exchange from retained vectors and position roles? First use a privileged diagnostic that subtracts the exact contribution of the other positions to check the target before choosing the learned design. Keep the vocabulary, codec and exchange fixed. Clean features would supply training targets only; evaluation would use new photographs and held-out position combinations, with corrected independent, untrained and shuffled-memory controls. This is a proposed experiment, not an established capability.