Selective Retention in a Bounded-Memory System: Sleep Consolidation When Importance Becomes Clear Later

Under a fixed memory budget, an experience can look unimportant when it occurs and turn out to matter only after its record has been discarded. We evaluate this problem in SLAB (Self-Learning Artificial Brain), a non-neural system that stores experiences as separate records. Before deciding which records to discard, SLAB uses later encounters and reactions to raise the priority of records that are still stored. In a scripted experiment with long-term storage sized for five days of simulated input, SLAB retained 91% of first-encounter records when evidence of their importance arrived three simulated days later, versus 2% with updates to previously consolidated records disabled (means across five seeds). At delays of one or two days, SLAB retained 100% of these records. The later encounters were enough for the system to know the items either way; the updates preserved the original encounters. On a separate eight-class drawing task with clutter, SLAB achieved 93% average final recognition accuracy, matching unlimited storage and exceeding oldest-first eviction (33%). On the lessons-only stream, where every record drew a reaction, it achieved 57% versus 95% with unlimited storage. Of four forgetting orders compared (most recent, most often met, reacted-to first, and the full score), only the full score held up on every test. Keeping whatever drew a reaction tied it wherever reactions alone marked what mattered, but fell to 36% on the lessons-only stream. The score also performed worse than random selection when used to choose neural-network replay examples on Split MNIST. In a daily-cycle simulation, background consolidation matched overnight consolidation on retention, with 99th-percentile event-handling latency of 0.37 ms versus 0.30 ms. These timings exclude perception and queueing and do not establish production availability. The results have not been independently reproduced. Code and raw results are available to researchers on request under a research-only license.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-20
DOI
https://doi.org/10.5281/zenodo.24173071
Primary Topic
Memory and Neural Mechanisms
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Selective Retention in a Bounded-Memory System: Sleep Consolidation When Importance Becomes Clear Later

Abdul Jaludi
Zenodo (CERN European Organization for Nuclear Research)
Memory and Neural Mechanisms
preprint

Selective Retention in a Bounded-Memory System: Sleep Consolidation When Importance Becomes Clear Later

Abdul Jaludi
preprint en

Abstract

Under a fixed memory budget, an experience can look unimportant when it occurs and turn out to matter only after its record has been discarded. We evaluate this problem in SLAB (Self-Learning Artificial Brain), a non-neural system that stores experiences as separate records. Before deciding which records to discard, SLAB uses later encounters and reactions to raise the priority of records that are still stored. In a scripted experiment with long-term storage sized for five days of simulated input, SLAB retained 91% of first-encounter records when evidence of their importance arrived three simulated days later, versus 2% with updates to previously consolidated records disabled (means across five seeds). At delays of one or two days, SLAB retained 100% of these records. The later encounters were enough for the system to know the items either way; the updates preserved the original encounters. On a separate eight-class drawing task with clutter, SLAB achieved 93% average final recognition accuracy, matching unlimited storage and exceeding oldest-first eviction (33%). On the lessons-only stream, where every record drew a reaction, it achieved 57% versus 95% with unlimited storage. Of four forgetting orders compared (most recent, most often met, reacted-to first, and the full score), only the full score held up on every test. Keeping whatever drew a reaction tied it wherever reactions alone marked what mattered, but fell to 36% on the lessons-only stream. The score also performed worse than random selection when used to choose neural-network replay examples on Split MNIST. In a daily-cycle simulation, background consolidation matched overnight consolidation on retention, with 99th-percentile event-handling latency of 0.37 ms versus 0.30 ms. These timings exclude perception and queueing and do not establish production availability. The results have not been independently reproduced. Code and raw results are available to researchers on request under a research-only license.

Zenodo (CERN European Organization for Nuclear Research)
Memory and Neural Mechanisms
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Selective Retention in a Bounded-Memory System: Sleep Consolidation When Importance Becomes Clear Later — Abdul Jaludi · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS