Faster visual recall by retaining reconstructed vectors

The question

Can reusing reconstructed VSA vectors reduce recall time without changing which memories are selected, what they contain, or what happens when new experiences arrive?

Why this experiment?

Once a recall path works, its repeated reconstruction and transfer costs can be measured without changing what counts as the right answer.

How it works

Repeat the same cue requests with and without a cache of reconstructed memory vectors. Check answers and rankings as well as latency and retained bytes.

What happened

Once populated, the cache reduced median search-and-recall-application time from 4.154 to 1.022 seconds: 75.4% lower. The 95th percentile fell from 4.679 to 1.109 seconds. Admitting a new photograph took a median 0.592 seconds without the cache and 0.590 seconds with it across eight paired admissions.

The tradeoff was 2.99 GiB of temporary GPU cache. The highest GPU tensor allocation measured across admission and query phases was 3.35 GiB, below the 6 GiB criterion. The cache is not saved with the memories and does not reduce durable storage.

All 384 paired decision comparisons agreed: 64 fields × three passes × two acceptance rules. Cached and uncached scores were exactly equal; checks of recalled content, remembered continuation, decoded images, new admissions and saved state also passed. These repeated comparisons come from 16 source photographs, not 384 independent inputs.

Every predefined correctness and resource criterion passed. This establishes a faster optional execution path on the tested history and hardware; it does not improve recall quality. The original unresolved queries and false match remained.

Limits and next question

Timing includes cue encoding, memory search and application of the baseline recall rule. Image decoding, remembered continuation, the alternative acceptance rule and persistence were checked separately, outside the timer. The 75.4% reduction therefore describes this query path, not the complete cognitive loop or application startup.

The first empty-cache query took 4.588 seconds. First queries after reload and clearing took 3.732 and 4.158 seconds, respectively; each is a single observation. A 64 MiB fallback cache had a 3.761-second median on an eight-field subset, which does not establish a full-panel speedup.

This is one device, memory history and process comparison. Cues were related views of 16 photographs. An unfamiliar cue with 64 visible features still produced a false recall under both implementations and both acceptance rules. The cache remains optional and is not enabled by default.

Next question. Return to acquired visual patterns and cleanup: can reusable patterns improve recall after vector operations without erasing distinctions between different experiences?